1762 lines
406 KiB
Markdown
1762 lines
406 KiB
Markdown
# Project State
|
||
|
||
## Review boundary repair — CPU validated, live 3x5 pending (2026-09-08)
|
||
|
||
- The latest actual Review Queue did not generate Chunk 1: backend history proved it reused all three saved chunks and returned the prior `complete` revision. The previous mode-selection presentation repair therefore did not satisfy the user's acceptance condition and is superseded by this entry.
|
||
- Source now treats an explicit Full Video -> Review Each Chunk selection as a new non-destructive review series. Before Queue it refreshes canonical Run Storage state; when saved `review_ready` or `complete` history exists, only the queued payload receives `Regenerate From = Chunk 1`, automatic nonce, and no Take selection. The visible settings and saved Takes remain unchanged. The first actual Sampler result releases setup mode and loads its canonical review state; later Continue queues inherit that new branch and advance one physical group at a time.
|
||
- Exact CPU contracts cover `3 x 5s`: fresh Chunk 1 -> review, Continue -> Chunk 2 review, Continue -> Chunk 3 complete. The real frontend lifecycle and backend branch/nonce resolver pass. Focused Review/UI suite `123` PASS and full CPU suite `1156/1156` PASS in `50.69s`; Registry/source declarations, JavaScript syntax and `git diff --check` pass. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-boundary-20260908`.
|
||
- Fixed H3情報チェック returned `PASS / local runtime deploy OK`; its nonblocking queued-identity note was applied by binding setup release to the exact setup token and Run name. Installed ComfyUI_W still has the superseded frontend SHA `d84877206cc1142dbfe7fce3f7e9c1b6edb8a22cae5ab117a1407f9435f6dea7`; validated source SHA is `fe5e00020b9e942ec2e27286576aca69fb3add3f18afeea8edcfd28f6183cd91`. Runtime deployment, browser refresh and actual GPU `3 x 5s` Chunk1/2/3 screen acceptance remain pending. No workflow, Production/Sampling backend, saved Take, commit, push, tag or release change.
|
||
|
||
## Local Review mode-selection repair (2026-09-08)
|
||
|
||
- Deployed locally after H3情報チェック returned PASS / local deployment OK (report-only review). Source, installed web/project_id.js and live8188 HTTP200 payload SHA all equal d84877206cc1142dbfe7fce3f7e9c1b6edb8a22cae5ab117a1407f9435f6dea7. Runtime snapshot verified1 file. Browser hard-refresh and user-visible check remain; no browser reload, GPU, backend restart or publication performed. This supersedes deployment-pending wording below.
|
||
- User-reported Full Video -> Review Each Chunk opened an old complete revision before generation. Reproduced with the real frontend harness (red.xml); mode selection now retains setup and the normal readiness summary without showing Retry/Restart. Existing Return to Review explicitly opens the saved result. Saved progress remains reusable; selecting Review does not mean starting a new Run.
|
||
- Presentation token is transient, not serialized or a backend identity. A matching queued generation's new review revision releases it; same-revision history reads, delayed initial history before Queue, a previously queued result, or another Run's result do not take focus. Hidden old Retry/Finish is normalized to ordinary Continue at Queue while manual boundaries remain intact. Reload recovery, review completion, saved takes, Python and workflow bytes remain unchanged.
|
||
- Focused90 PASS; final full1155 PASS in48.44s, CPU-only. Initial full had one old UI fixture failure: it injected a ready revision without Queue; fixture now models queued completion and keeps its prior expectations. Evidence: D:/Codex/_test_results/ComfyUI-H3-Continuum/v38-mode-settings-20260908. Completion report sent to fixed H3情報チェック; local deployment/readback pending that review. No commit/push/GPU or backend restart authorized/performed.
|
||
|
||
## Review UI hotfix — validated for user-authorized main publication (2026-09-08)
|
||
|
||
- User explicitly authorized a hotfix commit/push to existing origin/main for the confirmed Xarfai Review restart UI mismatch. Scope is frontend only: withdraw stale review actions after settings edits, preserve manual regeneration boundaries, add a non-destructive one-shot Start again from Chunk 1 button. Saved takes, sampler/transport/Audio/schema/workflow files and installed runtime remain unchanged.
|
||
- Browser-side edit detection only withdraws stale actions; it does not approve saved-prefix reuse. Existing Run Storage compatibility checks remain authoritative. The separate 15-second Review chunk2 static/noise report is unresolved and is explicitly excluded from the hotfix claim.
|
||
- Related pre-fix diagnostic evidence: D:/Codex/_test_results/ComfyUI-H3-Continuum/v38-xarfai-review-20260908/FINDINGS.md. New validation/publication evidence: D:/Codex/_test_results/ComfyUI-H3-Continuum/v38-review-hotfix-20260908. Initial source snapshots: nine verified one-file snapshots under before-review-hotfix-*. Final checks and publication result are recorded in WORKLOG and the task response.
|
||
- Final full guarded CPU suite1155/1155 PASS (46.62s), focused61 PASS, Registry105/105 and source declaration256/256 hashes match. All guarded processes exited CUDA-uninitialized, zero blocked initialization calls. Final additions cover delayed history response and queued-baseline isolation between Run Names. User authorized waiving the external review once due delayed delivery; the fixed H3情報チェック reply subsequently arrived with PASS / PUSH TO MAIN OK. Non-blocking review note: keep revision updated_utc tied to backend writes, not ordinary history reads. No GPU or Registry/tag/release-asset publication.
|
||
|
||
## Main publication preparation (2026-09-07)
|
||
|
||
- User explicitly approved commit/push to existing origin/main, then prioritized immediate README/file upload after announcing on Civitai and requested no further workflow tests. Publish the current Spectrum graph as MiniMax_H3_Continuum_V38.json and its matching ZIP, unchanged: JSON SHA e490c87e4995930f95c8fffc4c678111d7897e74003e1962b9b4a247ba43f4ac; ZIP SHA 83d28cc0cd549066962dc21ceae05d2b189fe22391b556887eab2cc1800e1e80. README, migration, package declarations and tests now describe one graph in two formats requiring Spectrum/rgthree/KJNodes/Easy Use. No separate Turbo graph and no prompt/media/title cleanup.
|
||
- English and Japanese READMEs link JSON/ZIP and LightX2V near the top with git clone/pull instructions. Git exact-byte attributes prevent automatic EOL conversion from invalidating file hashes. Historical V3.7 ZIP remains a source regression fixture; unrelated node.zip and external backup archives are not publication payload.
|
||
- Publication audit initially ran 1,155 tests: 1,154 passed; the real-Core CPU cache subprocess stopped at optional SageAttention import requesting CUDA capability. Initialization was blocked. The CPU cache harness now models that optional package as absent in its child process only; installed Core/Sage and Production remain unchanged. Focused rerun is recorded in v38-main-publication-20260907/focused.xml. Do not relabel the initial run as full-suite PASS. No further workflow test runs after the user's stop instruction.
|
||
- Evidence: D:/Codex/_test_results/ComfyUI-H3-Continuum/v38-main-publication-20260907. Source snapshots verified before changes. This is GitHub source publication only, not a GitHub Release/tag or Registry publish. Runtime synchronization and Registry official validate/pack remain separate; user backend and GPU were untouched.
|
||
|
||
## README accuracy corrections (2026-09-07)
|
||
|
||
- Corrected the English/Japanese README only: insert Core Create Video before Save Video; distinguish automatic nonce0, explicit positive nonce, interrupted resume and review Retry; scope historical compatibility promises to their releases and preserve the exact current seven-node boundary; identify Audio Switch as Core; align Japanese Advanced/Manual visibility guidance; separate the latest CPU1,154/1,154 audit from historical V3.6/GPU acceptance.
|
||
- Refreshed only README-related hashes in Registry (104/104 matched) and source Manifest (254/254 matched). Final focused CPU regressions90/90 PASS; initialization tripwires stayed clean. README local links53/53 resolve and Markdown fences/whitespace checks pass. Evidence: D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-readme-correction-20260907. Initial six one-file snapshots verified under before-readme-correction-*-20260907_195135.
|
||
- No workflow, image, ZIP, generation code, public schema, installed runtime, backend, commit, push or release change. The prior runtime synchronization remains pending separately; documentation checks do not promote release/GPU-quality status.
|
||
|
||
## CPU release hygiene fixes and ZIP scope correction (2026-09-07)
|
||
|
||
- The previously inspected ZIP was D:\Codex\_git_push_work\ComfyUI-H3-Continuum\node.zip, not the user's Continuum.zip / Continuum3-8.zip backups. Its distribution role was never established; withdraw the ZIP-based launch blocker. Backup/local ZIPs are excluded from this release check and remain unchanged. The older audit report carries this correction at its top.
|
||
- Fixed only hardening.format_memory_snapshot's CUDA observation guard: read counters only after CUDA is already initialized. Added three CPU tests for no probe/no tensor or RNG mutation, initialized-counter preservation, and nonfatal diagnostic failure. The real-Core cache test now uses explicit CPU bootstrap, initialization tripwires, and the diagnostic-only Kitchen capability bypass.
|
||
- Full guarded pytest: 1,154/1,154 PASS in55.60s, zero deselection and zero intercepted initialization calls in recorded guard logs. Focused tests60/60 PASS. Registry104/104 and source Manifest254/254 refreshed for current declared files. No Sampling, Audio, mask, seed, SIGMAS, public schema, workflow, or ZIP change.
|
||
- Reconstructed105-file Registry payload (104 declared files plus manifest) under D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-cpu-release-fix-20260907\registry_payload, without using or creating a ZIP. Isolated CPU import, public7-node registration, PackedLayout, diagnostics and atomic JSON replacement PASS; CUDA remains uninitialized.
|
||
- Runtime update plan is a dependency-complete two-file sync of hardening.py and run_storage.py; only added import is stdlib time. One-file snapshots of each installed target verified. Deployment attempted no writes because user queue was running1/pending0. Existing backend remains untouched; disk synchronization is pending an idle queue, and an eventual user-managed restart is still needed for the running process to load the changes.
|
||
|
||
## V3.8 prelaunch CPU audit (2026-09-07, Release HOLD)
|
||
|
||
- Fresh guarded full suite: 1,149 passed / 2 failed, 1,151 total in 99 modules; no deselection. Failures: stale Registry manifest and real-Core cache test launching without CPU bootstrap. Same cache harness passes all four assertions under explicit CPU/W2 bootstrap; original tests remain unchanged.
|
||
- Source CPU/Core 0.34.5 probe: all seven public nodes, callable inputs, native PackedLayout and Fixed prompt self-test PASS. V3.8 core workflow 19 nodes/25 links and user Spectrum 32 nodes/38 links pass graph/schema checks; external Spectrum/KJ/Easy/rgthree sockets checked via GET metadata only. Spectrum remains byte-identical to user Downloads SHA E490C87E4995930F95C8FFFC4C678111D7897E74003E1962B9B4A247BA43F4AC. Python226/JS7 syntax PASS.
|
||
- Historical observations: Registry manifest4/source Manifest7 mismatches before record update (subsequently repaired above). Root node.zip has75 files and an incomplete import, but is not an established distribution artifact and is excluded from the launch verdict. Runtime source diff was functional only in run_storage.py (source-only Windows atomic replacement retry); reference.py and v3/reference_video.py differences are whitespace/line endings with matching AST.
|
||
- New independent CPU diagnostic finding: hardening.format_memory_snapshot reaches CUDA lazy init through mem_get_info even with Core --cpu when CUDA is available. Full suite blocked14 such calls plus1 Core startup call; standalone CPU diagnostic reproduces. All actual initializations were blocked. No Production/workflow/runtime fix, model load, Sampling, Queue submission, backend restart, commit, push, or publication.
|
||
- Source README local links exist; 44 references across shipped docs point to non-Registry assets and need GitHub/local-documentation handling. Full report and evidence: D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-prelaunch-cpu-audit-20260907\CPU_PRELAUNCH_REPORT_JA.md. Record-only snapshots before-cpu-prelaunch-audit-state/worklog-20260907_182042 verified1+1 files. CPU wiring PASS does not grant GPU-quality or Release PASS.
|
||
|
||
## V3.8 Civitai feature infographic (2026-09-07)
|
||
|
||
- Rebuilt the supplied V3.7 feature infographic as a V3.8 Civitai graphic at the exact original `1536x1024` 3:2 landscape format. The new content covers Review Each Chunk, Resume/Regenerate, simplified V3.8 controls, and the shared Spectrum/Turbo graph; it explicitly retains the V3.7 Production foundation and states that Issue #13 remains open. Saved as `docs/images/v38-manual/h3-continuum-v38-whats-new.png` (SHA-256 `5E4B876585FAC3A2D580CDAF99B83EC3E10DA90036B7923B4DFFDCF0240560BD`). No workflow/runtime change.
|
||
|
||
## Spectrum-only optional V3.8 workflow and top-level Git instructions (2026-09-07)
|
||
|
||
- Added the user-supplied `MiniMax_H3_Continuum_V38_Spectrum.json` to `examples/workflows/` without changing its contents. Source and repository SHA-256 are both `E490C87E4995930F95C8FFFC4C678111D7897E74003E1962B9B4A247BA43F4AC`; its 32 nodes / 38 links parse with zero missing endpoint nodes.
|
||
- Kept `MiniMax_H3_Continuum_V38.json` as the only dependency-free Registry workflow. The Spectrum graph is a GitHub-source optional workflow and is excluded by the existing Registry packaging rules. No separate Turbo JSON is published; README explains how the same optional graph can switch from its saved Spectrum/20-step state to one task-matched LightX2V Turbo LoRA after Spectrum is disabled.
|
||
- Added `git clone` and `git pull --ff-only origin main` near the top of the English README and linked the LightX2V/ModelTC GitHub plus Hugging Face project. Per user direction, the Spectrum workflow's empty prompt, media filenames, node titles, widgets, and links were not edited.
|
||
- Validation: focused release/public workflow tests `11 passed`, with only the existing Registry manifest hash test deselected pending the final package freeze. Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-spectrum-readme-workflow-20260907_162216`. No commit/push/tag/release.
|
||
|
||
## English README complete V3.8 UI reference draft (2026-09-07)
|
||
|
||
- Expanded the English README from the Review-only walkthrough into a complete reference for every user-facing socket/control/output on the seven-node V3.8 public surface. It now distinguishes Main, Advanced, Review/Take, three Continuum Settings, loader bypass, First/Last/Reference Image roles, Video Guide, Driving/Reference Audio, Reference Audios, Finalize, and Second Pass. Hidden frontend-managed IDs are not presented as editable UI.
|
||
- Added a top-level provider-neutral “ask an AI to read this manual” path using current official ChatGPT/Gemini/Grok information. The text does not promise that a pasted GitHub URL always imports a repository; Gemini's explicit Import code route and README paste/upload fallback are stated.
|
||
- Copied seven user-supplied PNGs unchanged into `docs/images/v38-manual/`: loader bypass, three Reference Images, First/Last Image, First Image Draft sizing, LightX2V LoRA selector, First Image tooltip, and Manual Width/Height. Source/destination SHA-256 matched for all seven; README now has 32 local image links and zero missing targets.
|
||
- Confirmed from source that Manual accepts 32x32, size selection is built into the V3.8 Sampler, loader OFF uses native bypass behavior, and the public surface remains exactly seven nodes. LightX2V's own Hugging Face/GitHub records were used to separate FL2VA/Ref2VA and 4/8-step targets; a 4-step LoRA at six inference steps is labeled workflow-specific rather than an upstream default.
|
||
- Recorded Issue #13 conservatively: still Open; Standard V3.8 has no active mitigation; the isolated 6x5 R6-A improvement did not become a Production claim because the later 4x8 gate produced gains 1.0 and A/C bit-exact no-op. This documentation change does not alter Production, runtime, workflows, manifests, Sampling, or backend state.
|
||
- Validation: focused tests `120 passed`; one registry-manifest test remains expectedly red because README plus pre-existing `run_storage.py` and `web/project_id.js` are not yet re-frozen into `REGISTRY_MANIFEST.sha256`. README image-link audit is `32/32`; no manifest was rewritten before the final workflow/package freeze. Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-v38-full-ui-readme-20260907_152253`. No commit/push/tag/release.
|
||
|
||
## English illustrated continuation manual live revalidation (2026-09-07, Continue PASS)
|
||
|
||
- Used the user-owned `127.0.0.6:8188` Chrome workflow `MiniMax_H3_00058_` through Computer Use with Turbo4 v1.2, actual Steps 4, Euler/simple, Spectrum OFF, Draft 0.30 MP, fixed seed, Review Each Chunk, Progress ON, and 2x5 seconds. The backend was not stopped or replaced.
|
||
- Fresh Chunk 1 reached the three-action review panel. Selecting `Use it and continue` did not execute by itself; the top-right blue `Run` queued the action. The next Queue sampled only Chunk 2 (Sampler UI time 43.738s), reused the accepted Chunk 1, and reached `Saved sequence is complete` at 10 seconds. First Queue Sampler UI time was 103.802s; do not interpret the cold/warm difference as a general speed ratio.
|
||
- Added exact UI crops for review actions, selected Continue, completed sequence, Back-to-Settings completion state, Render History actions, and Regenerate From with visible Variation Nonce. README removed the obsolete retry-failure warning, uses the new screenshots, explains that Chunks is the final total, and updates the diagnostic four-step expectation to `sampling_steps_total=4`.
|
||
- Evidence full screenshots are under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-manual-live-20260907`. Snapshot before edits: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-manual-steps4-capture-20260907_142631`. No commit, push, tag, release, or backend shutdown was performed.
|
||
|
||
## V3.8 Continuation Regression Final Gate (2026-09-07, Pre-GPU PASS / approval pending)
|
||
|
||
- Supersedes the stale restart-required review status below. The repaired Review state was live-validated: `Try this chunk again` regenerated the selected chunk and returned the three-action review screen; after the user disabled an accidentally enabled Spectrum node, `Continue / Next` completed normally. Manual screenshots remain deferred.
|
||
- Current task follows the supplied `H3 Continuum V3.8 — Continuation Regression Final Gate`: compare current Standard/Balanced22/Audio ON Masked AV against Compatibility Reference Context under one frozen `4×8s`, `896×896` GPU A/B. No Production transport change is allowed unless GPU result is Case A.
|
||
- Snapshot `pre-v38-continuation-regression-final-gate-rg-20260907_110054`:935 entries, manifest SHA `DD62DCA87437CABB049D16BFB720DF45FAC04988096C7A3827433D95206CD698`, creation-time source mismatch0. Normal snapshot helper hit the inherited pytest-cache ACL; verified robocopy+rg fallback retained the same authoritative source.
|
||
- Code audit confirms V3.4 final code baseline `d3dab7b` uses fresh-target Reference Context; current Standard/Balanced22/Audio ON uses masked AV target prefixes and Compatibility uses Reference Context. Both remove the later-group hard First Image keyframe while retaining original visual/Qwen identity tokens.
|
||
- Actual Comfy CPU Shadow PASS: Standard has exact Video7/Audio37 tail prefixes with zero prefix masks and no continuation refs; Compatibility has one exact AV ref, fresh zero target, no mask, target-head MM-RoPE alignment and preserved `position_ids` identity. CUDA remained uninitialized; owned PIDs/8190/HTTP ended.
|
||
- Pre-GPU focused5/5 and syntax PASS; stage/source runtime153/153 exact. Frozen semantic graph diff is only `continuation_backend`; Spectrum/Sol/Still/Driving/Reference/Last/Run Storage/Second Pass/Hi-Res/R6-A are OFF. Protocol SHA `7c525971a0a98694f75915b304c349b5f95d59ae0d450a07efd98ceb91ab4736`; execution manifest SHA `3f801370f69e8709f24a6849535483fbb00d49978ba419acf779c127d7afd412`.
|
||
- GPU authorization artifact is absent. Negative launch test refused before creating a run or opening8190. GPU Sampling, A/B result, reset/seam and Case A/B/C are pending explicit approval. User-owned8188 is untouched; no Production/commit/push/tag/release.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-continuation-regression-final-gate-20260907\PRE_GPU_REPORT.md`. The previously recorded fixed-task id was no longer readable; the current task list resolves title `H3情報チェック` to ChatGPT conversation `6a8ef934-10ac-83e8-9af2-9e2417af7224`, but detailed handoff was rejected because this changed destination identity lacks fresh explicit authorization. No workaround was attempted. Mandatory review and GPU remain pending user authorization.
|
||
|
||
## Regenerate From / provenance reuse repair (2026-09-07, installed / restart required)
|
||
|
||
- Immediate current task: reach the user's actual Chunk1-review screen with Continue / Retry / Finish buttons, then STOP before generating Chunk2. Manual screenshots and transport research remain deferred. Do not fake readiness or delete existing media.
|
||
- Actual UI run29e95cfc-1c9d-41fd-b32c-0d07855353a6 submitted Review Each Chunk,2x5s,Regenerate From Chunk1,nonce0,Continue / Next,Take Automatic. Unexpected result: cb575f0c6aa00705 nonce1/new_revision,2 reused/0generated,10s video00070,52.35s. This is NOT successful first-chunk regeneration.
|
||
- Root cause: prepare's provenance candidate path checked ancestry/lineage/prompt but ignored changed per-chunk sampling contracts. `_valid_provenance_prefix` now optionally enforces actual producer-revision chunk-contract hashes from the requested branch boundary onward; prepare enables it. Prefix before the cut remains reusable, same-nonce interrupted work remains resumable, explicit Take selection remains separate. Stored-manifest validation is unchanged. No new hard stop or Sampling/Audio/Assembly/schema change.
|
||
- Also repaired blank Render History: actual custom canvas widget now draws selected/canonical Take text instead of relying on an unsupported multiline primitive. Existing transient serialization guard retained. Lifecycle fixture verifies that Selected text is actually drawn.
|
||
- Tests107 PASS (9 new on-disk provenance cases plus review/branch/control/execution suites); UI19 PASS; extra canvas-draw fixture PASS; diff check PASS. First test collection failed due to importlib-mode fixture import and was corrected without production changes. Stage AST confirms ONLY prepare/_valid_provenance_prefix differ from previous runtime, exactly matching tested source, zero import changes; old source-only atomic JSON helper remains excluded.
|
||
- Installed runtime run_storage SHA0E402AD912926AE02FDC2A50BED9009B4C3A090EF6D52F6344C3B26FA33C1676; JS SHA7B318BE7E0A6FFA7A0B924F149EE90C10E4EBFB731DA2BC40297BAD0ECF63269. Full runtime snapshot181files before-review-reuse-runtime-20260907_104350; source scoped4files before-review-reuse-scoped-20260907_104350 (normal helper hit inherited cache ACL).
|
||
- User-owned8188 remains running old Python code, queue0/0 at deployment. User restart and browser reload REQUIRED before another Run. Then keep Regenerate From Chunk1 for exactly one Review execution, verify1generated/1of2available,review_ready and visible3buttons. Reset manual Regenerate From to Auto through UI before later continuation (do not rerun Chunk1 inadvertently). Smart Retry GPU still unverified; no completion handoff/commit/push/release. Previous full-suite2 failures remain documented below.
|
||
|
||
## Completed-review UI repair (2026-09-07, deployed / browser reload and GPU retry pending)
|
||
|
||
- Supersedes the older diagnostic-only status below. Review identity repair is installed in runtime run_storage.py (SHA59593daf4e2021bad2e6849427aae95a439220745f582b406ea1d240b19ebe45): canonical set observation and Core archived parameter dtype observation prevent false LoRA/VAE identity changes. CPU169 focused tests passed. Post-fix fresh Chunk1 `45dc7891-f3a8-4995-8bd7-25c1f3e959d2` and Continue `ffff7252-33b9-4f38-8bf7-fe3033c2d453` succeeded; revision99684d4a78b5855b, Chunk1 file SHA unchanged, only Chunk2 generated, 10-second video00065. Smart retry GPU remains unverified.
|
||
- Current UI failure independently reproduced: `/view?...project.json` returns200, canonical99684d4a78b5855b is complete with review_unit2, while frontend reviewReady accepted only review_ready. Backend already permits Regenerate Current for complete+review_unit. User Runs638b84ae and2c52c22f actually submitted Continue / Next and reused2/generated0, not retry; latestvideo00068, workflow25.34s.
|
||
- Changed web/project_id.js only for the runtime UI: backend-derived completed panel remains accessible; valid reviewed unit allows retry; completed Continue/Finish hidden; history/settings retained even for complete without a retry unit. No Sampling/Assembly/schema/status/persisted-revision changes. Source/runtime JS exact SHAB571FD4E061B113931D5470DD9844CAD77390B5227CC0EF05576ECCDA9662812. Scoped web snapshots7+7files and records3files verified under before-completed-review-*20260907_101839.
|
||
- New real-JS lifecycle fixture covers initial/partial/running/interrupted/completed,1..6 chunks, Terminal Merge label, one-shot queued retry, settings round-trip serialization parity, history, reload and failed fetch. Focused128 passed; additional frontend15 passed with one distribution-manifest hash failure. Full suite:1139 passed /2 failed (same manifest mismatch and identity test assumes process-global CUDA remains uninitialized after earlier tests; isolated identity suite passes). Do not claim full-suite PASS or pure CPU-only full-suite execution. Package/Release remains HOLD; these two failures need separate disposition.
|
||
- Auto-review blocked browser Ctrl+R because current00058 tab is dirty (`*`). No reload or extra GPU Queue occurred. Ask the user to save/reload or explicitly approve reloading the unsaved tab; do not circumvent the block. Then verify completed UI, select Try this chunk again, submit once, compare saved Chunk1 SHA and generated/reused counters. User-owned8188 remains running. Manual screenshots and Standard/Compatibility proposal remain deferred; no commit/push. Completion handoff is pending the live repair gate, not yet complete.
|
||
|
||
## Review retry diagnosis follow-up (2026-09-07, restart required / root cause still unresolved)
|
||
|
||
- User authorized fixing the live retry failure. Read-only CPU load of saved revision `52e4e5124afd5d7a` passes `_validated_review_head` and `find_latest_review_head` against its stored contract: prefix1, review_ready, no notes, CUDA uninitialized before/after. Saved data corruption is not demonstrated; the actual retry-time contract remains uncaptured. A model/runtime fingerprint change is a hypothesis, not a confirmed cause.
|
||
- Added hash-only mismatch diagnostics to `run_storage.py`: when no compatible head exists, report up to three nearest stored contracts and differing contract sections. Retry retains its original error and rejection policy; no unsafe fallback, identity relaxation, Sampling/seed/mask change, or automatic fresh generation. Successful lookup skips the new comparison. Tests: Run Storage/review79 plus policy/cache/UI/branch83 =162 PASS; diff check PASS.
|
||
- Installed only the three diagnostic-definition changes into runtime `run_storage.py`, SHA `16f0bb17b863dc0b1aef95656deae9a4783338b35145f010bec95a661feae6df`. AST audit proves no other runtime/import changes and exact agreement with tested source for those definitions. The source's pre-existing atomic-replacement retry helper was intentionally NOT included in this narrow deployment. Full runtime snapshot181 files: `before-review-mismatch-runtime-20260907_093306`; source scoped snapshot4 files/SHA4: `before-review-mismatch-scoped-20260907_092947` (normal source helper encountered existing cache ACL).
|
||
- Diagnostic code is on disk but the user-owned8188 process has NOT been restarted or stopped; queue0/0. User restart is required before further Computer Use retry diagnosis. No new GPU run, publication, final manual validation or completion handoff. Resume using evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-manual-live-20260907\RETRY_DIAGNOSTIC_STATUS.md`; do not claim the underlying bug fixed.
|
||
|
||
## English illustrated continuation manual (2026-09-07, Computer Use recovered / retry validation FAIL)
|
||
|
||
- Computer Use now initializes successfully. Used the existing Chrome tab `927290933` (`00058`, user-owned `127.0.0.6:8188`) directly. Exported the current UI workflow before edits and snapshotted five files at `before-live-manual-scoped-20260907_090852`. Changed only seed control to fixed and Run to Review Each Chunk (UI enabled Progress automatically); retained supplied Turbo4 v1.2 / actual4 Euler/simple / Spectrum ON / KJ Sage / blank prompt / 2x5s / effective544x544.
|
||
- First UI Run PASS: prompt `e2c6038a-6403-4b27-bc4e-4d390a61741e`, backend execution75.232s, sampling30.447705s,120frames/5s, video `D:\output\video\comfy_video\video\MiniMax_H3_00064_.mp4`. Review panel actually appeared. New exact-pixel crops show settings, blue top-right Run, review-ready, and selected retry; README distinguishes top-right execution Run from the in-node Run setting (review messages still say Queue).
|
||
- Second UI Run, Try this chunk again: FAIL before new sampling, prompt `f8c18378-7a42-40b5-8408-19e8e86b877b`, `Regenerate Current requires an existing reviewed unit`. Full submitted graph differs only by `review_action`. Saved manifest remains `review_ready`, unit1..1, resume_safe=true; saved chunk SHA matches manifest. The read-only trace locates missing compatible review-head resolution, but why it rejects/misses this stored head remains unresolved. Do not blame a seed change or claim retry/continue/finish verified. Stop further GPU runs pending diagnosis/fix authority.
|
||
- Runtime/source check: review_control.py, v3/nodes.py, web/project_id.js exact; run_storage.py differs only by the source's Windows atomic JSON replacement retry helper. No evidence links that difference to the observed missing-head error. No runtime installation or code fix performed. Queue running0/pending0; user backend intentionally left running. Source README has an explicit draft-validation warning; no commit/push.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-manual-live-20260907` contains exported workflow, both full prompt/history records, summaries, saved manifest, full screenshots and UI action timestamps. Failure screenshot kept as internal evidence, not public manual artwork. Remaining history/settings/continue/finish/regeneration/extension screenshots and final release documentation review are pending. Earlier tool-blocked notes below are historical.
|
||
|
||
- README now contains a general 11-section Review/continue/retry/finish/settings/history/regeneration/extension/timing guide, with six unmodified user-supplied UI crops under `docs/images/v38-manual/`. Two chunks is only an example; total-vs-added count, fixed Base Seed, later-section rebuild, and Last Image/Terminal Merge exceptions are explicit. README_JA is unchanged by this task.
|
||
- Source/code cross-check and six PNG relative links plus section anchors PASS; no claim of fresh UI or arbitrary extension GPU validation. Prior two-Queue six-step timing is identified separately; downstream time is not individually profiled. GitHub-ready source assets only; Registry/package refresh and commit/push are not performed.
|
||
- User then supplied a running ComfyUI at `http://127.0.0.6:8188/`, workflow `00058`, Turbo at four steps, and requested live Computer Use screenshots plus execution timing. Use that user-owned backend after visual verification; do not stop or replace it. No Queue, UI edit, or new GPU execution has occurred in this documentation task.
|
||
- Both `node_repl` (`@oai/sky` bootstrap) and `cua_repl` fail before execution with `failed to write kernel assets: 指定されたパスが見つかりません。 (os error 3)`. Resetting each kernel did not resolve it; workspace/TEMP/TMP exist. Exact missing internal path and root cause are unknown. No unsupported screenshot/desktop automation fallback was used. Live capture remains blocked pending tool recovery; record new observed UI before any action.
|
||
- Verified four-file pre-edit snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\before-illustrated-review-readme-scoped-20260907_085726`; standard helper first failed on inherited `.pytest_cache` ACL. Resume from README and the current-scope note in `docs/V38_BEGINNER_MANUAL_PLAN_AND_CAPTURE_SCRIPT_JA.md`. Completion review is pending until live documentation validation is complete; do not send to the unpinned duplicate H3 task.
|
||
|
||
## Issue13 R6-A Turbo4 v1.2 / actual 6-step / 0.30 MP A/B (2026-09-07, VALID / NEGATIVE-HOLD)
|
||
|
||
- Executed the supplied Turbo4 v1.2 / 6-step Euler/simple / Draft0.30 MP settings as an adapted `6x5s` efficacy test, not an exact reproduction of the supplied `2x5s` workflow. Effective size was `544x544`; prompt remained empty; Audio Continuity/Balanced22/KJ Sage were ON. Fresh Group1 was shared; A had R6 OFF and B differed only by frozen global latent AC contraction.
|
||
- Execution contract PASS: both arms are `720f/24fps/30s`, stereo32k PCM `960000` samples; Assembly exact; Sampling traces complete; resources `5/5` with hard-stop/contract errors zero; both dedicated backends normal-shutdown with owned PIDs gone,8190 closed and HTTP offline. Exactly five policy evaluations occurred: G2 required no-op, G3 natural no-op, then gains G4 `.9`, G5 `.95440975`, G6 `.9`. Raw AV is exact through G3 and diverges from the first effective intervention at G4.
|
||
- Numeric promotion evidence is NEGATIVE/HOLD for this condition: endpoint Laplacian `+8.28%`, Gradient `+11.85%`, Contrast `+113.44%`, Shadow `+71.07%` worse; motion proxies about `1.6%` lower and seam metrics warn worse. Highlight endpoint improved `99.14%` and mean-saturation endpoint improved `20.20%`, but they do not offset the primary/guard negatives. Perceptual A/B playback remains pending; seam ratios are not a perceptual-cut claim.
|
||
- Interpretation is narrow. The empty prompt produced severe shared semantic excursions before R6 acted, including near-white G3 and green-background G5, so this is weak evidence for Issue13 natural cumulative drift. Do not generalize to R6 overall or invalidate the separate earlier Turbo8 numeric candidate. No Production/public schema/default/package or release change.
|
||
- The first attempt stopped pre-Sampling on expected-plan evidence ordering; it was preserved and the isolated runner ordering was corrected without graph/Sampling change. The completed wrapper later exited only because atomic `RESULT.json` refused to overwrite that first attempt; all A/B artifacts were durable and shutdown complete, so no GPU rerun was needed.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-turbo4-6step-0p3mp-20260907\TURBO4_6STEP_0P3MP_REPORT.md`. Fixed `H3情報チェック` completion review `bd47c1b1-8aa6-4f87-b24a-efa7c8d74fe7` accepted `EXPERIMENT VALID / EXECUTION CONTRACT PASS / NUMERIC PROMOTION EVIDENCE NEGATIVE / HOLD / DO NOT PROMOTE` with no mandatory correction. Source-record snapshot `pre-r6-turbo4-report-record-filtered-20260907_081550` has `919` files after the standard helper met the inherited `.pytest_cache` ACL. No commit/push/tag/release.
|
||
|
||
## Issue13 R6-A resumed B: numeric candidate, perceptual review pending (2026-09-07)
|
||
|
||
- Fixed completion summary review25ee8534/9747f470 read: accepted with explicit NUMERIC GPU EXPERIMENTAL CANDIDATE qualifier, not perceptual PASS. Full30s A/B playback/listening is next, before additional GPU/new workflow experiments. Review did not independently inspect local media/artifacts. Final source230 code hashes unchanged; diff --check0; B41648/40316/36912 absent,8190 listeners0.
|
||
- User explicitly resumed B before supplied Turbo/Spectrum workflows or Audio Continuity experiments. Completed A evidence reverified (R3B rawAV12/12 plus exact historical decoded/PCM and owned shutdown); interrupted B retained. New fresh B0b0b607953934a6bb2e7ed7142784ae6 completed13m08s from shared savedG1 under unchanged frozen protocol79bbe9a5..., after53/53 resume CPU tests and fixed-task prelaunch review542d0853/3bfd0ff3 accepted inside-lock checks. No numerical method or threshold changed.
|
||
- EXPERIMENT VALID / FROZEN NUMERIC GATES PASS / PERCEPTUAL REVIEW PENDING. Lap endpoint drift39.86749->13.36942 (66.47% reduction), Gradient72.91%,Highlight91.17%,Contrast26.82% better. Saturation4.10% worse; seam1.38317->1.51302 (9.39% worse); both within frozen guards. Motion/detail/PCM numeric screens PASS. These are metric-drift changes, not overall quality percentages.
|
||
- Counterevidence: largest benefit is Group6; G5 Lap nearA and Gradient slightly worse. Peak Lap excursion reduction only15.15%, so uniform long-trajectory suppression is NOT established. Paired stills are coherent but full playback/listening and perceptual identity/detail/seam PASS remain unverified. One scene/seed only, not Issue13 solved or Production promotion.
|
||
- Actual invariants/noise5/5 match, rawG1/G2AV4/4 exact, scope violations0,720frames/30s/960000sample stereo32kHz PCM. GainsG2..6=1,.991727,.995747,.994723,.973105. Both owned backends normal shutdown/PID/8190/HTTP confirmed; user backend untouched. No additional GPU phase started.
|
||
- Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pre-gpu-20260907\FINAL_REPORT.md`, QUALITY_SCREEN.json,PAIR_REVIEW. Two supplied workflows retained unchanged in user_workflows, NOT executed/mixed intoB. Completion review requested at fixed H3情報チェック before any next phase. Source record snapshot919 pre-r6-B-completion-record-excluding-cache-20260907_063034. Production/Release/A5/general A8b HOLD; no commit/push/tag/release. Earlier user-paused state below is historical.
|
||
|
||
## Issue13 R6-A: Standard A complete, B USER-PAUSED (2026-09-07)
|
||
|
||
- Latest user instruction stopped continuation and requested analysis of A only. GPU authorization is now false; do not resume B or start another GPU run without a new user instruction. The stop arrived after B Group2 had begun (last log2/8 steps); immediate launcher interrupt ended the run. This is not a B quality FAIL or completed experiment.
|
||
- Fresh A completed13m04s: R3B rawAV12/12 exact, historical decoded metrics and PCM exact, five Sampling/noise records, intervention-scope violations0, normal A shutdown confirmed. Review of the corrected PreGPU gate was accepted by fixed summary reply70782869-06e2-461a-96fb-0edb978f4eea before launch.
|
||
- All six A/B owned PIDs41304/21224/20080/19020/2276/4140 independently confirmed gone;8190 listener0/HTTPoff. B normal queue-drained shutdown was not established because immediate interruption took priority. User backend untouched. Pause evidence `issue13-r6a-pre-gpu-20260907\USER_PAUSE.json`.
|
||
- A-only CPU analysis:720f/24fps/576x864/30s, central25-frame means show Lap248.55->288.42 (+39.8675 units, +16.04%, NOT +39.87%), Gradient+5.44%, highlight clipped fraction0.1794%->0.5662%. Sampled subject/composition remains coherent; late edge/texture strengthening persists. Strongest normalized seam at15.0833s=1.3832. PCM intact/exact but RMS~-54.9dBFS; no listening/full-playback perceptual PASS claimed. This is baseline reproduction, not mitigation success.
|
||
- Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pre-gpu-20260907\A_ANALYSIS\REPORT.md`, summary/seam/detail/join artifacts. Fixed summary review c4a59d67/f1b2f17a accepted interpretation, emphasizing that observed late-stage edge strengthening is not a fresh causal proof, nor a playback/listening PASS. Review does not authorize GPU resume. Source-record snapshot919 `pre-r6-A-analysis-record-excluding-cache-20260907_034446`. Production, commits, push, release unchanged.
|
||
|
||
## Issue13 R6-A corrected Pre-GPU freeze (2026-09-07, supersedes prior freeze)
|
||
|
||
- User approved correction, CPU verification, refreeze and the single Fresh A -> conditional Fresh B GPU pair; no renewed GPU permission is needed for that pair. Fixed-task completion review is pending before launch. GPU efficacy is still untested.
|
||
- Independent local inspection corrected signed-negative Laplacian acceptance, seam threshold discrepancy, unrelated two-secondary-improvements requirement and missing actual Sampling invariant recording. Absolute Laplacian threshold remains31.89399101047766; per-chunk Lap/Gradient>=80% shared Group1 is an added blur screen, not perceptual proof. Secondary absolute drift<=1.2A, motion>=.8A, seam<=1.2A. One-seed candidate only.
|
||
- Symmetric observer records actual seed/noise/SIGMAS/conditioning/ordered wrapper and sampler evidence; B must match current A before model execution. Intervention source/reference/Audio/generated-region/mask hashes remain exact; Group2 prefix no-op. Completion manifest requires both sampling_trace and pin_renorm evidence. Fresh A rawAV12/12 and exact historical decoded/PCM parity remain mandatory.
|
||
- CPU46+controller/lifecycle45=91 PASS; CUDA tripwire clear in correction suite; canonical/stage230 code SHA exact; upstream14 and execution259 hashes exact. New protocol SHA79bbe9a5b671ed2d6d2c66d0d549e759ab28bef1b4a67938cd0e80a3049a442d. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pre-gpu-20260907\CORRECTION_GATE_REPORT.md`.
|
||
- Research pre-edit snapshot353 files; pre-record source snapshot919 files `pre-r6-correction-record-excluding-cache-20260907_031923` after inherited .pytest_cache ACL prevented ordinary helper completion. Production/commit/push/release unchanged. No automatic retry or sweep.
|
||
|
||
## Issue13 R6-A Pre-GPU Gate PASS (2026-09-07)
|
||
|
||
- Frozen comparison is one pair only: A is Standard/Balanced22 with the R6 hook disabled; B differs only by enabling the global latent AC scalar contraction on protected Video-prefix slots. Shared saved Group1, 576x864/24fps/6x5, seed131325, FL2VA INT8/Turbo8/Euler/simple/KJ Sage and all other settings are fixed. This is not a texture- or frequency-specific correction.
|
||
- Fresh A must pass R3B raw Video6/6 + Audio6/6 exact plus PCM/artifact/resource/Assembly/owned-shutdown checks. It must also exactly reproduce frozen historical A decoded metrics and PCM after excluding only run ID and output path. Any raw or decoded baseline difference prohibits B. Historical Laplacian drift39.8674887631 and B threshold<=31.8939910105 remain fixed; motion>=80%A, seam<=1.2xA, secondary Video metrics, PCM and perceptual review are guards.
|
||
- Dedicated CPU tests passed55, including all existing fail-closed cases, authorization refusal, graph parity, and exact-A-quality acceptance plus a 1e-9 motion-metric rejection. Upstream hashes14/14 and execution files251 are frozen. Protocol SHA `48d110db7d376622f4275125207220a997d3134c1ecdf48276cd37d65fab936a`. GPU authorization artifact is absent; no backend/GPU ran;8190 listener zero. Production remains unchanged.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pre-gpu-20260907\PRE_GPU_REPORT.md`. Snapshot `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-r6a-pregpu-record-excluding-cache-20260907_024434` has919 files; ordinary helper first stopped only on inherited `.pytest_cache` ACL. Fixed review aa0cb846-a49f-44ea-a3e3-e682d6a144d5/9a9dc5e1-3363-459c-bc3e-86511dfb97e2 required decoded-A parity; after correction, review91ee90f4-ebf4-4659-a329-e85747c4aac0/a5c3e54b-6ff7-4a08-84f2-a11c58a42664 accepted final Pre-GPU PASS with no remaining design correction. GPU still requires explicit immediate approval.
|
||
|
||
## Issue13 R6-A actual Comfy CPU Shadow PASS (2026-09-07)
|
||
|
||
- One owned CPU-only Comfy backend completed an actual Queue/history request using staged public V3.8 registration, Core NestedTensor, the real masked AV prefix builder, and saved R3B rawAV. No model/VAE/noise/Sampling. OFF returned original latent object for Groups2..6. ON Group2 gain1.0 was exact no-op; only Groups3..6 protected Video first7 slots changed with gains0.9917270/0.9927173/0.9882362/0.9733208.
|
||
- Generated Video, whole Audio object/content, masks object/content, metadata, conditioning fixture, SIGMAS, current tails, fixed Group1 reference, saved raw12 and Python/Torch RNG remained exact. Five events shared one reference SHA.9 prototype tests and15 actual verification checksPASS.
|
||
- CUDA false at import/bypass/node/shutdown; no tripwire. OwnedPIDs9704/36796/29444 ended; queue empty,8190 listener/HTTP off; user backend untouched. Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-actual-comfy-shadow-20260907\REPORT.md`. Snapshot `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-r6a-shadow-record-20260907_013958` has924 files.
|
||
- Pre-GPU Gate remains incomplete: actual model Sampling OFF rawAV12/12, noise/derived-seed/full-conditioning/wrapper/Assembly invariants, execution hash freeze, and decoded AV quality remain GPU-stage evidence. GPU not started and requires explicit approval immediately before launch. User-authorized fixed summary review 6ae817e0-7654-4c06-a06b-75bd9eee42fc /a1bab86c-09b8-4132-aa30-91e2f85bb637 accepted classification with no mandatory correction. R6 does not directly change Sampling-input Audio, but joint AV generation may indirectly change output Audio; GPU must measure it. Review is not independent local-artifact inspection.
|
||
|
||
## Issue13 R6-A Pin Renorm CPU design audit PASS WITH CONDITIONS (2026-09-07)
|
||
|
||
- Isolated non-Production prototype freezes Group1 final seven Video slots as the Balanced22 reference and applies `current_mean + gain*(current_tail-current_mean)` only to the protected seven Video prefix slots for Groups2 onward. `gain=clamp(reference_std/current_std,0.90,1.00)`; amplification prohibited. Audio37T, masks, generated regions, seed/noise/SIGMAS/conditioning/Assembly/State/Session/RunStorage remain unchanged by contract. This is global latent AC scaling, not texture-only/frequency-selective.
|
||
- Actual R3B shadow requested gains feeding Groups2..6:1.0000000,0.9917270,0.9927173,0.9882362,0.9733208. Thus Group2 is unchanged and intervention first appears at Group3; maximum measured reduction2.67%. Per-channel ratios are heterogeneous, so decoded efficacy/motion safety remain unproven.
|
||
- Focused CPU9 passed: OFF object identity and RNG neutrality, Group1 exclusion, source/reference immutability, Video-prefix-only scope, Audio/mask/metadata preservation, bounded/no-amplification gain, and invalid/degenerate/shape/transport fail handling. Actual rawAV finite/expected shapes; CUDA remained uninitialized; no backend/GPU Sampling. Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r6a-pin-renorm-cpu-20260907\REPORT.md`.
|
||
- Valid pre-record snapshot `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-r6a-cpu-record-20260907_012541` has924 files. GPU not authorized by this CPU Gate; next phase requires isolated actual-Comfy OFF parity/Shadow/lineage wiring and frozen decoded evaluation, followed by explicit GPU approval under the latest instruction. User-authorized fixed summary review 05b4095a-89d0-4262-9323-d5932f3ea630 /f8e60c80-b3d1-4637-9cbb-48f490d178f6 accepted PASS WITH CONDITIONS. Numeric degeneracy contract is explicitly CPU float64,epsilon1e-12; reference <=epsilon fails, current <=epsilon no-op. Review is not independent local-artifact inspection.
|
||
|
||
## Issue13 R5 5f actual GPU pair closed: PRIMARY QUALITY FAIL (2026-09-07)
|
||
|
||
- Actual fresh A matched R3B raw Video6/6 + Audio6/6 bit-exact. PCM/artifact/plan/resource checks and full owned-backend shutdown passed before fresh B. B applied immutable Group1 first5f reference exactly for Groups2..6. Both720frames/30s and stereo32kHz960000sample PCM. Execution/integrity PASS; no OOM/nonfinite.
|
||
- Primary Laplacian drift A+39.86749 -> B+74.97015 (88.05% larger drift; not overall image-quality percentage), failing frozen <=31.894. Gradient drift91.44% worse. Several tone/color drifts improved, but cannot override primary failure. Motion/seam numerical guards PASS; PCM numeric screen clean, listening/full-playback perceptual verdict not established.
|
||
- DO NOT PROMOTE this fixed5f compound content+rows+AV-position intervention. No sweep; Issue13 unresolved; Release HOLD. No Production code/commit/push/tag/release change. Source230 py/js hashes match924-file pre-run snapshot. A/B owned processes zero and8190 listeners zero rechecked after completion; user backend untouched.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-gpu-ab-20260907\FINAL_REPORT.md`. Fixed H3 information-check summary review f62e27e4-6b3c-477b-aa0b-7a25c9d20454 /827767df-4345-48e2-b511-f36b1eeeb583 read: no mandatory correction; not independent artifact review.
|
||
|
||
## Issue13 R5 GPU A/B authorized; preparation frozen (2026-09-07)
|
||
|
||
- User expressly authorizes continued GPU testing without repeated confirmation, within current R5 single5f A/B scope. Previous GPU-unapproved entries are historical. A actualR3B rawAV12/12 +PCM/artifacts/resource/plan/normalshutdown remain mandatory beforeB. No sweep/Production/publication authority added.
|
||
- Isolated runner/graphs at `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-gpu-ab-20260907`; PRE_GPU_REPORT.md and final execution hashes.14upstream/assets/protocol hashes matched;4CPU preparationtestsPASS. GPU not started at this record; mandatory fixed review requested before launch. DirectGPU startup never uses W2 CPU bypass.
|
||
|
||
## Issue13 R5-W2 scoped CPU wiring PASS (2026-09-07)
|
||
|
||
- Fixed summary review c03e9c5b-823c-4f74-93e1-bba5a8bd48c2 /56226d50-cf39-4acf-ab2c-1f5949fa947f read; accepts scoped PASS, no mandatory correction. Synthetic evidence is not quality/GPU parity. Actual A12/12 and R5 ON execution remain POST-approval GPU checks; preparation/freeze remains CPU work, not authority to start GPU.
|
||
|
||
- CPU bootstrap alone bypasses kitchen INT8 availability; lazy/C initializer tripwires retained, CUDA availability API unmodified. Actual Fresh Comfy A/B fixture Queue/history and PCM-observed stage Finalize completed720frames/960000samples each with observer parity/RNG unchanged and CUDA uninitialized through preexit. No model/VAE/Sampling; R5 source and GPU Sage/KJ unchanged.
|
||
- Real A live denied B; current A synthetic raw12/PCM/evidence verified only after queue0/0 and owned PID/listener/HTTP shutdown, then rechecked under exclusive lock before CPU fixture B. Raw12 fixture is NOT GPU baseline parity. A41480/2120/9176 and B24760/5132/30700 ended. 15 offline evidence assertions plus37 regressions PASS. Source230 unchanged vs snapshot924 pre-r5-w2-20260907_000922.
|
||
- Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-w2-cpu-bootstrap\REPORT.md`. Full R5-2 NOT YET PASS: actual GPU runner quality metrics/resource/plan bindings and final execution freeze remain. GPU unapproved. Fixed summary review requested. No Production/publication change.
|
||
|
||
## Issue13 R5-W1 first CUDA initialization cause isolated (2026-09-07)
|
||
|
||
- Fixed-task summary review read:738aad5c-f71f-43fa-af16-b0f82de4a8e1 /0938013b-95fc-4e41-963b-0ff48cddec72. Accepts first-cause isolation PASS only; no remediation, no exhaustive-trigger claim, no inference of normal GPU Sage defect. Next CPU-specific workaround requires its own scope; not started.
|
||
|
||
- One guarded startup intercepted Core attention.py54 -> comfy_kitchen.sage_attention.is_available:75 -> torch.cuda.get_device_capability -> _lazy_init before actual initialization. torch initialized false before/after; intentional exit86. No prompt/model/Sampling or automatic retry. First observed cause is CPU-unconditional Core/kitchen capability probe, before R5 registration; later triggers not ruled out.
|
||
- Nine offline checks PASS; source230 py/js unchanged against snapshot924 pre-r5-w1-init-isolation-20260906_235631. Owned26156/33568/15832 ended,8190 listener/HTTP offline. Evidence `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-w1-init-isolation\REPORT.md` plus full stack, command, result and hashes.
|
||
- Cause isolation PASS only; no remediation or installed-runtime edit. Full R5-2 NOT YET PASS; GPU unapproved. CPU-specific availability-probe avoidance remains proposed, retaining the tripwire and preserving eventual GPU Sage configuration. Fixed-task review requested.
|
||
|
||
## Issue13 R5 actual Comfy wiring HOLD (2026-09-06)
|
||
|
||
- Isolated stage/Sampler/PCM Finalize graph adapters created; nine CPU tests PASS. Actual Comfy object_info and nonce-only prompt/history binding worked. No model or Sampling request submitted.
|
||
- Strict probe detected torch.cuda.is_initialized=True despite --cpu and Device:cpu. CPU-only integration is NOT PASS; initializer unresolved. Do not automatically repeat startup or weaken the guard. All owned PIDs4812/10092/38456 and earlier attempts gone; queue0/0,8190 listener/HTTP offline. Production unchanged.
|
||
- Full R5-2 NOT YET PASS: real metrics/resource/plan binding, controller integration and final freeze remain incomplete. GPU unapproved. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-comfy-wire-20260906\REPORT.md`. Snapshot pre-r5-comfy-wire-20260906_233825:924 files. Fixed-task summary review requested; awaiting reply.
|
||
|
||
- Fixed review ec956664-22e1-4846-8023-e574b7967a2b / 02c95f2c-ef0e-42d4-83a9-afd26c748f4e read: summary-only PARTIAL wiring PASS, CPU-only claim rejected, full NOT YET PASS. CUDA runtime initialization is observed; GPU computation is not established by that fact. Next is initialization isolation and full controller/evidence/freeze, not GPU sampling.
|
||
|
||
## Issue13 R5-2 CPU launch lifecycle (2026-09-06)
|
||
|
||
- CPU lifecycle sub-Gate PASS: 8 tests, actual minimal loopback CPU service A/B start/shutdown, PID tree/create_time/8190/HTTP verified stopped. No ComfyUI/model/torch/CUDA/GPU. Cross-process exclusive lock, fresh controller identities/root, atomic completion-last publication with bounded retry, immutable decision hash and B recheck verified. First test exposed Windows read-only fsync Errno9; diagnostic r+b flush correction passes, no B launched on failure and A cleaned up.
|
||
- Final A PIDs14284/18348/38780 and B25704/40888/39560 exited;8190 absent. Fixture12/12 is not GPU baseline parity. Source230 unchanged versus snapshot924 pre-r5-launch-lifecycle-20260906_231926, diff check0. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-launch-lifecycle-20260906\REPORT.md` and JSON/XML/frozen evidence.
|
||
- Fixed summary review `fd337f43-4b58-494a-bbf6-6339635fb0d7` / `913264ed-3b1b-42f2-8aaa-d8acd33930dc` accepts CPU Launch Lifecycle PASS. Full R5-2 NOT YET PASS pending actual ComfyUI prompt/history/resource/shutdown adapters and real R5 stage graph wiring/freeze. Future actual GPU A12/12 is post-approval B permission gate, not a prerequisite to running CPU preparation. GPU remains unapproved; no publication.
|
||
|
||
## Issue13 R5-2 CPU evidence evaluator / Finalize adapter (2026-09-06)
|
||
|
||
- Sub-Gate PASS: 37 CPU/Shadow tests. Evaluator reads actual safetensors keys/shape/dtype/SHA, recomputes PCM slices/assembly/pad/capture, checks artifact completeness/hashes/run provenance/settings/plan/resource and live-OS probe interface. Tests mock OS/launch only; no backend/GPU. Actual historical R3B 12 raw files hashed, 14 CPU assets frozen with mismatch0; source230 unchanged against snapshot924 pre-r5-runner-control-20260906_230811.
|
||
- Fixed review turn `fa12ae03-03e4-430e-a2eb-26408a82ab1b`, reply `cde10501-5542-4b83-b075-864125170228` accepts sub-Gate PASS only (summary review). Full Pre-GPU remains NOT YET PASS pending real launcher/controller context ownership, exclusive lifecycle, atomic completion-last publication, and actual shutdown E2E. Review proposes non-sampling service for shutdown E2E, but none was started; current CPU-only/no-backend scope is preserved and any service-start authority must be resolved before execution.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-runner-control-20260906\REPORT.md`, SHADOW_TESTS.xml, R3B_BASELINE_RAW.json, HASH_FREEZE.json. No Production/runtime change, GPU approval, commit or push. GPU observer Sampling transparency remains unproven until separately approved Fresh A actual12/12.
|
||
|
||
## Issue13 R5-2 real Assembly/Finalize PCM integration (2026-09-06)
|
||
|
||
- CPU integration sub-Gate PASS: real Finalize/Assembly with synthetic decoded stereo32k PCM and 8x8 images, 85 tests (14 integration/predicate +25 detector +46 Assembly regression). No VAE/model/backend; CUDA uninitialized. Source230 code files unchanged versus snapshot924 `pre-r5-pcm-integration-20260906_224938`.
|
||
- Complete sample ledger checks input/copy order, exact retained content, hidden padding, natural958667 -> repeat-tail1333 -> final960000, capture SHA and restoration/RNG/input parity. Shortfall hidden by native padding, duplicate copy, wrong final pad and wrong capture point are detected. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-pcm-integration-20260906\REPORT.md`, JSON/XML/fixtures and 23-file hash freeze.
|
||
- Fixed-task review delivered and read: turn `26c56b98-8172-4f3c-bd58-ff3d391d851e`, reply `771b11bf-9d7f-49f1-88c9-bdcd8d16335a`. Summary-only review accepts sub-Gate PASS. Full R5-2 remains INCOMPLETE pending runner integration, real file-hash A12/12 + actual shutdown E2E control, and final runner/prompt/prototype/observer execution-hash freeze. No GPU approval requested; user explicitly requires GPU waiting.
|
||
|
||
## Issue13 R5-2 Audio CPU preparation (2026-09-06)
|
||
|
||
- Audio CPU sub-Gate PASS: offline NumPy PCM screen and capture helper, 25 tests passed. Stereo32k/960000 samples, actual five cuts, per-channel RMS/spectral/edge/click/dropout limits frozen. Numeric clean means review pending; confirmed Audio harm fails R5 regardless of Video improvement.
|
||
- Natural Assembly 719f/958667 samples -> exact 720f/960000 adds 1333 samples. Future isolated runner must capture decoded counts, retained slices and pre/post adjustment; final count alone cannot establish no hidden shortfall.
|
||
- Full R5-2 remains INCOMPLETE: executable runner/capture integration, CPU validation and runnable prompt/observer hash freeze remain pending. No GPU/backend or Production edits. Model/media/shared Group1 hashes 8/8 match; 244 CPU assets frozen; source230 code files unchanged from snapshot924. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-2-pre-gpu-20260906\R5_2_AUDIO_CPU_REPORT_JA.md`.
|
||
- Following explicit user CPU sharing permission, the previously denied report was delivered: turn `b738494e-a202-4736-a9c8-a2c6380d110c`, reply `bf27b4da-9d2c-4313-95dc-4d43b1f04867`. Review accepted Audio CPU PASS and thresholds, requiring real-path PCM ledger/capture integration. This requirement is now CPU-tested as recorded above. GPU requires separate later approval.
|
||
|
||
## Issue13 R5 position-control CPU audit (2026-09-06)
|
||
|
||
- Same-layout zero Reference is constructible, but it is not a neutral no-anchor control: Core default augmentation makes all 93,312 zero-latent row values nonzero; removing augmentation does not remove projection/modulation/unmasked-attention participation. Metadata without latent leaves 972 required condition rows missing.
|
||
- Within the current native Reference contract and without row/attention/position rewrites, an inert matched control was not found. Reclassify Standard-versus-R5 as generated Group 1 reference plus additional conditioning rows and Core-native AV positional shift; do not claim pure appearance-content causality or general mathematical impossibility of controls.
|
||
- Zero-ref versus real-ref remains a possible non-neutral content ablation, not a no-anchor baseline. No extra GPU arm has been authorized or run. R5-1 PASS WITH CONDITIONS remains; Pre-GPU design uses 5f only, Run Storage OFF and explicit shared saved Group 1.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-position-control-20260906\POSITION_CONTROL_REPORT_JA.md` and `POSITION_CONTROL_RESULT.json`. CPU assertions passed, source tensor immutable, CUDA uninitialized, backend/GPU Sampling absent. Fixed-task turn `57811644-8575-4e2d-87cd-930a7065ad45`, reply `5ac513f9-ba7e-4814-87c4-d42bff3f2518` accepted `Position-Matched Control Audit PASS` and the limited reclassification. Next freeze is Standard A versus R5 B only, with Video quality, motion/identity, Video join, Audio join and Audio continuity guards. GPU remains unapproved.
|
||
|
||
## Issue13 R5-1 isolated CPU/Shadow prototype (2026-09-06)
|
||
|
||
- Implemented Default-OFF diagnostic scope, immutable 5f Group 1 Video-only reference, ordinary/Terminal sequence integration, exception cleanup and ON identity separation only in `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-1-prototype-20260906\prototype_src`. Production 230 Python/JavaScript files match the 924-file pre-change snapshot.
|
||
- CPU tests: 88 passed; actual unmodified Core PackedLayout definitions: 10 cases passed. OFF sequence parity, Group 2-6 fixed anchor, user-reference presentation/tags, Replay/Review/Branch and Terminal shadows passed. These are stub-sampling structural tests, not GPU quality or kernel parity.
|
||
- Conditional finding: native reference addition shifts absolute target Video and Audio positions by 8.333333 while preserving Audio values/count/relative timing. Do not claim absolute Audio position invariance. Persistent old-lineage selection after Group 1 replacement and ON disk-restart E2E remain unimplemented; proposed GPU scope must use explicit saved Group 1 and Run Storage OFF.
|
||
- Two full-Core import attempts initialized CUDA as an import side effect and exited without model loading/forward/Sampling/backend. Final layout runner executes unmodified Core layout AST definitions and verifies CUDA uninitialized. Ports 8188/8190 have no listener. Release/A5/general A8b HOLD; no publication.
|
||
- Formal verdict: CPU/Shadow PASS WITH CONDITIONS. Fixed-task turn `87d13aed-685c-4b5c-b13b-22a484da8766`, reply `f64a95b6-219d-4f8c-9073-006290220483` accepted it and requires a position-matched control feasibility audit before R5-2, or explicit reclassification as anchor plus native AV positional shift. A zero latent is not assumed neutral. No custom Core position correction is allowed. Evidence: `R5_1_REPORT_JA.md`, `R5_1_TESTS.xml`, `R5_1_PACKED_ROWS.json`, `R5_1_SOURCE_PARITY.json`. GPU remains unapproved.
|
||
|
||
## Issue13 R5-0 Dual-Role Anchor / Tail CPU architecture audit (2026-09-06, PASS WITH CONDITIONS)
|
||
|
||
- Audited a long-term Group 1 anchor and the existing short-term previous-tail transport without starting a backend or GPU. The current I2VA Standard path already retains the original First Image as Qwen visual tokens while Balanced 22 uses the immediately previous `22f Video / 37T Audio` as a masked target prefix; R5 therefore must add only the generated Group 1 appearance cue, not duplicate the existing first-image role.
|
||
- Rejected a Group 1 frame-zero keyframe: Core `PackedLayout` places it at the same position as the masked target head, so it competes directly with the previous-tail pose. The viable prototype route is one fixed Group 1 head `5f / 2 video slots` as a Video-only `minimax_refs` block, appended after existing user refs, while the previous `22f / 37T` AV tail, masks, restoration, trim, and Audio ownership stay unchanged.
|
||
- CPU evidence using the frozen R4 Standard Group 1 raw Video proved source immutability and separate segments. Baseline rows were `21,501`; the 5f Video-only reference produced `22,473` (`+972 / +4.52%`). A 22f reference is structurally valid but deferred at `24,903` rows (`+15.82%`) because it repeats substantially more Group 1 motion and approaches the R2.6 over-constraint mode.
|
||
- Shadow `PackedLayout` also accepted the hidden Video-only anchor after existing image/audio refs while retaining a final-frame keyframe; hidden anchor Audio rows remained zero. This proves packability only. Quality improvement, motion freedom, user-reference semantics, wrapper runtime behavior, and GPU cost remain unproven.
|
||
- R5-1 may proceed only as an isolated Default-OFF diagnostic policy: OFF bit-exact, Group 1 and tail immutable, one anchor from Group 2 onward, no Audio in the anchor, existing reference order preserved, no new schema, and safe Standard fallback when only an initial State is available without the Group 1 entry. Review/branch/replay and Terminal Merge must pass CPU/Shadow before any GPU proposal.
|
||
- Fixed-task review turn `3f6fc599-b9b0-4ca9-8e5e-8910cce205d6`, reply `3179d562-3414-41bd-8e38-d91ec4bf03a5` accepted `PASS WITH CONDITIONS` and froze three R5-1 prerequisites: hidden anchor is appended after all user refs without changing numbering/presentation/tags; PackedLayout row ranges and positions for target/current-tail/user-refs/hidden-anchor/Audio are measured separately without custom row-order or position-ID rewriting; and the anchor remains the immutable Group 1 of the run lineage across replay, branch, and Review.
|
||
- Formal verdict: `R5-0 CPU ARCHITECTURE AUDIT = PASS WITH CONDITIONS`. Production, public schema, Sampling v5, seed/SIGMAS, State, Session, Run Storage v3, Assembly, Release, A5, and general A8b remain unchanged/HOLD. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r5-dual-role-anchor-audit-20260906\R5_DUAL_ROLE_ANCHOR_AUDIT_JA.md` and `R5_CPU_AUDIT_RESULT.json`.
|
||
|
||
## Issue13 R4 Native Guided Overlap final GPU A/B (2026-09-06, VALID / FAIL / DO NOT PROMOTE)
|
||
|
||
- After explicit GPU approval and the user's durable resource-policy correction, executed the frozen full pair in the required order. Fresh Standard A completed `6x5` and reproduced the R3B raw Video `6/6` plus Audio `6/6` exactly (`12/12`), then its Queue `0/0` backend was fully stopped before Fresh R4 B started.
|
||
- R4 B completed `720` frames / `30.000s` Video and `30.000s` Audio with `6` finite raw AV groups and no OOM, allocation error, backend crash, NaN/Inf, packed-layout, AV-alignment, or duration hard stop. Hidden overlap did not extend final duration.
|
||
- Formal verdict: `ISSUE13 R4 NATIVE GUIDED OVERLAP = GPU A/B VALID / FAIL / DO NOT PROMOTE`. Laplacian drift worsened from `+39.8675` to `+74.1887` (`+86.1%`), Video join was `2.1225 > 1.6598`, Audio join was `3.3896 > 1.6069`, highlight drift worsened `282.9%`, and gradient drift worsened `74.2%`. Saturation improved `63.0%` and the frozen motion floor passed, but they do not offset the primary failures.
|
||
- B sampling was `22.0%` slower and peak backend RSS was `47.24 GiB` versus A `43.29 GiB`. Per user instruction, low free VRAM and RSS alone are advisory for future approved GPU runs; actual OOM/allocation/backend/integrity failure is the stop boundary.
|
||
- Fixed task `H3情報チェック` completion review turn `a2e7fa47-e1ec-44c4-b83c-ff9e9872f25a`, reply `e7884bfc-6a46-4aad-8618-3ebead295311` accepted the verdict without mandatory correction. The result is limited to this frozen one-scene, one-seed condition; it does not invalidate Native Guided Overlap in general.
|
||
- Do not retry, sweep, or promote R4. The next research candidate may be a CPU/read-only design audit of `Group 1 persistent anchor + current tail`, keeping Video and Audio intervention roles separate and preserving clean tail, State/Session/Assembly, Replay/Terminal/Review, and Core layout contracts.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-final-ab-20260906\R4_FINAL_AB_REPORT_JA.md`, `R4_FINAL_AB_RESULT.json`, `A_RAW_PARITY.json`, and the preserved A/B raw AV, videos, resource logs, decoded analyses, contact sheets, and boundary sheets.
|
||
- A owned PIDs `40224/36876` and B `9164/41744` were each stopped after Queue `0/0`; port `8190` listener and HTTP are offline. User `8188` was absent and never operated. Production/runtime/contracts and commit/push/tag/release remain unchanged.
|
||
|
||
## Issue13 Standard Group 2 repeatability (2026-09-06, PASS / prior M3A divergence not reproduced)
|
||
|
||
- After fresh explicit GPU approval, ran the frozen Standard Group 2 trace on three fully independent fresh 8190 backends. All three reproduced the R3B generated Video SHA `38e8eb8f...a188` and Audio SHA `2b482cfc...12d8` exactly.
|
||
- Every run matched all `150/150` ordered trace boundaries and all `8/8` predict-noise, diffusion-forward, and sampling-step records. Different/missing were `0/0`, first divergence was `null`, runtime/model options/wrapper order matched, and hook transparency passed `3/3`.
|
||
- Reverified the frozen model, CLIP, Video VAE, Audio VAE, Turbo8 LoRA, First Image, Core H3 nodes/model, KJ Sage patch, and shared Group 1 AV hashes. OOM, NaN/Inf, and unexpected backend crash were `0`.
|
||
- Formal verdict: `STANDARD GROUP 2 REPEATABILITY PASS / PRIOR M3A DIVERGENCE NOT REPRODUCED`. This closes the Group 2 repeatability blocker but does not identify the earlier M3A cause and does not assess R4 efficacy.
|
||
- Fixed task `H3情報チェック` completion review turn `0da950e6-2fc3-4c75-8a69-9daac82c9374`, reply `1d3c4727-9259-4832-8219-ae227a9477b1` accepted the verdict without mandatory correction. The next authorized proposal must resume the original frozen Gate order: fresh Standard A full `6x5`, require R3B raw AV `12/12` exact and no hard stop, stop A backend, then and only then run one frozen R4 B in another fresh backend.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-standard-group2-repeatability-20260906\STANDARD_GROUP2_REPEATABILITY_REPORT_JA.md` and `STANDARD_GROUP2_REPEATABILITY_RESULT.json`; raw traces remain under the prior M3B result root with `standard_group2_repeat_1` through `_3` case names.
|
||
- Run 1 PIDs `18612/39884`, Run 2 `19416/39548`, and Run 3 `6968/4652` were each stopped after Queue `0/0`; 8190 listener and HTTP are offline. User 8188 was absent and not operated. Production/runtime/contracts and commit/push/tag/release remain unchanged. R4 B and final video were not run.
|
||
|
||
## Issue13 R4-M3B Group 2 GPU trace (2026-09-06, PASS / no divergence reproduced)
|
||
|
||
- After explicit GPU approval, executed the frozen Group-2-only diagnostic in the mandatory order. The R3B control generated-region hashes reproduced historical Video `38e8eb8f...a188` and Audio `2b482cfc...12d8` exactly, so and only so the R4-M3A Standard harness ran in a separate fresh backend.
|
||
- The two current traces are bit-exact through all 150 ordered boundaries: conditioning, pre-sampler latent/SIGMAS/noise/mask, eight `predict_noise` calls, all eight diffusion forwards, all sampling-step `x/x0`, sampler output and Group 2 generated Video/Audio. Different/missing are `0/0`; model options, wrapper order, memory runtime and PyTorch/CUDA settings also match.
|
||
- The first raw comparator mismatch was only the isolated Python module location (`stage_r3b` versus `stage_r4`) for the same `model_patch.py` SHA and callable. Comparator post-processing now ignores only that capture-location string when source SHA/name/qualname match; capture contract remains frozen at SHA `39affe50...19c2`. Dedicated tests pass `13`.
|
||
- Formal verdict: `R4-M3B GPU TRACE PASS / NO DIVERGENCE REPRODUCED / PRIOR M3A CAUSE UNRESOLVED`. The earlier R4-M3A output mismatch is not reproducible under this one frozen pair, so no internal first-divergence point or cause can be assigned. This one pair does not prove repeated determinism.
|
||
- R4 Native Guided Overlap B, quality A/B, retry and sweep were not run. R4 quality remains unassessed and requires a new direction decision before any additional GPU work.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-m3b-first-divergence-20260906\R4_M3B_GPU_RESULT_JA.md`, machine result `R4_M3B_GPU_RESULT.json`, ordered comparison `R4_M3B_FIRST_DIVERGENCE.json` and both raw traces/logs.
|
||
- Fixed task `H3情報チェック` completion review turn `d9289f52-2ed6-49f5-937a-73ca63e6efe3`, reply `ebb530f8-372d-4707-8636-84d3dc306980` accepted the verdict with no mandatory correction. It explicitly forbids calling this one-pair match `baseline recovery confirmed` and prioritizes a separate fresh-backend Standard Group-2 repeatability Gate before any R4 Native Guided Overlap quality A/B.
|
||
- R3B backend PIDs `7608/35932` and R4 PIDs `35984/39160` were stopped after Queue `0/0`; port 8190 listener and HTTP are offline. User 8188 was absent and not operated. Production/runtime/contracts and commit/push/tag/release remain unchanged.
|
||
|
||
## Issue13 R4-M3A baseline parity root-cause audit (2026-09-06, first contract divergence identified / output cause unresolved)
|
||
|
||
- CPU/read-only comparison of R3B baseline, the earlier R4 A `retry2`, and R4-M3 A `rerun2` found the first contract-level bit-exact break at Assembly Plan `chunks[0].context_frames`: R3B `5`, R4 shared-Group-1 harness `0`.
|
||
- This metadata difference is not the Group 2 output cause: R4 `retry2` has the same value `0` and nevertheless matches every R3B Group 1-6 Video/Audio tensor bit-exact. `entry_to_state()` does not consume the plan's `context_frames` field.
|
||
- The first saved tensor SHA divergence is Group 2's generated region: Video begins at latent `T=7`, Audio at `T=37`. The protected 7-slot/37-tick prefixes match the exact Group 1 tail and remain bit-exact.
|
||
- Normalized prompt graph, Group 2 derived seed, shape, scheduler settings and CPU-reconstructed target/mask/noise recipe match. R3B did not capture actual conditioning, actual SIGMAS, pre-sampler latent, first-forward, or per-step SHA, so A/B/C internal classification remains unresolved.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-m3a-parity-audit-20260906\R4_M3A_BASELINE_PARITY_ROOT_CAUSE_AUDIT_JA.md` and `R4_M3A_FIRST_DIVERGENCE_AUDIT.json`. No GPU, backend, Production change, A rerun, B execution, commit, push, tag, or release.
|
||
- After separate explicit user authorization, the fixed task `H3情報チェック` reviewed this CPU audit (turn `b9306a6b-b2da-486e-80d8-0b054eb80f76`, reply `7e67d0bf-783b-4a52-a2d7-d8d9cb2f6c16`). It accepted the formal verdict and the R4 B STOP. The next Group-2-only instrumentation must also capture model options, wrapper registration/order, attention patch identity and Dynamic RAM/runtime settings. The `context_frames=0` difference is not sufficient by itself to explain the output divergence; interactions are not excluded.
|
||
- R4 quality remains unassessed. Any next in-sampling telemetry run requires a separately frozen diagnostic plan and fresh user approval immediately before GPU execution.
|
||
|
||
## Issue13 R4-M3 GPU rerun (2026-09-06, INCOMPLETE / A baseline parity stop)
|
||
|
||
- After explicit GPU approval, corrected only the isolated observer runtime hook from the facade `v3.driving_nodes` to the actual Production binding `v2.nodes.run_sequence`; compile/static checks and R4 focused CPU tests passed `4/4`. Production source and installed runtime were unchanged.
|
||
- Fresh-backend A `A_standard_6x5_s131325_rerun2` completed Group 2-6 Sampling and produced 720 frames / exact 30.000 seconds / 24 fps / 576x864 / 32 kHz stereo in `780.011 s`, with no OOM, allocation error, NaN/Inf, backend crash, layout, AV, or duration failure.
|
||
- The mandatory pre-B check failed: frozen Group 1 Video/Audio remained exact, but Groups 2-6 Video/Audio did not reproduce the prior R3B baseline tensor bit-exact (`2/12` exact total). Derived seeds, prompt and geometry remained fixed; this is not a file-metadata-only difference. Per the frozen protocol, B and all retries/sweeps were not started, so R4 quality and the Group 5 B-A memory delta remain unassessed.
|
||
- A Group 5 RSS was `41.715 GiB` before Sampling, `41.854 GiB` immediately after, and `41.846 GiB` after cleanup. Overall minimum available RAM was `2.989 GiB`, peak commit `72.219%`; no hard resource stop fired.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-native-guided-overlap-20260906\R4_M3_GPU_RERUN_RESULT_JA.md`. Dedicated 8190 Queue was `0/0` before stopping listener PID `34336` and launcher PID `34940`; HTTP/listener shutdown was verified. 8188 was not operated. No commit/push/tag/release.
|
||
- Fixed-task completion review turn `c476af69-2f07-4dfb-a495-fea7f881199e`, reply `7259d478-7029-4077-9e8c-94d94f83ce78` confirmed the verdict with no mandatory correction. Next is a CPU/read-only-first bit-exact reproducibility cause audit; no further GPU A/B is authorized by this result.
|
||
|
||
## Issue13 R4-M1 Memory Ownership Audit (2026-09-06, PASS / GPU rerun awaiting approval)
|
||
|
||
- CPU/read-only ownership audit found no R4-specific major leak or unbounded Group-to-Group retention. The R4 guide clones are call-local, share one tensor pair across conditioning entries, and are released with conditioning; fresh targets are not referenced by R4 conditioning. Existing full physical CPU entries and the latest continuation tail are bounded A/B-common ownership required by the current Assembly Plan.
|
||
- Prior R4-3 wording is corrected to `R4-3 INCOMPLETE — SOFT RESOURCE THRESHOLD EXCEEDED`. Groups 2-5 sampled normally; the old 46 GiB exception fired before the Group 5 result returned to sequence cleanup. R4 remains a continuation candidate, not HOLD / FAIL.
|
||
- CPU verification: weak-reference ownership check PASS; existing R4 6-Group and Terminal Merge shadow tests `4 passed`; modified isolated files compile. At 576x864 BF16 the R4-specific 22f guide clone is about 0.623 MiB video plus 0.0045 MiB audio, not a GiB-scale explanation for backend RSS.
|
||
- The isolated R4 observer and test copy now treat RSS >=46 GiB as a Soft Warning, record Sampling-before/after and Group-cleanup snapshots plus system RAM/pagefile/commit/GPU/time, and defer resource stops until cleanup where possible. Production and installed runtime remain unchanged.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-native-guided-overlap-20260906\R4_M1_MEMORY_OWNERSHIP_AUDIT_JA.md`. The unchanged frozen R4-3 A/B GPU rerun is eligible only after fresh explicit user approval immediately before execution.
|
||
- Fixed-task review was supplied/read: `R4-M1 PASS`, no mandatory correction, and Pre-GPU conditions may be presented. The required follow-up is an explicit A/B Group 5 comparison at Sampling-immediate and post-cleanup points.
|
||
- R4-M2 Pre-GPU final freeze is `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r4-native-guided-overlap-20260906\R4_M2_PRE_GPU_FINAL_CHECK_JA.md` with machine freeze `R4_M2_PRE_GPU_FINAL_CHECK.json` SHA-256 `b31daaf81c00534df8811755d47b956ee27c16b34e00772722427fdaaef3c959`.
|
||
- Fixed-task review turn `72fc3888-2e94-4911-a531-39598e855576`, reply `13c5d2dc-0013-436c-8d23-3c5bafd48ab0`: R4-M2 PASS WITH CONDITIONS. The one mandatory correction is applied: backend health failure now has mechanical process/API/stall criteria. Results must emphasize fresh-process run-internal RSS deltas. This review is not GPU authorization.
|
||
|
||
## V3.8 beginner manual plan and capture script (2026-09-06, draft ready / capture deferred)
|
||
|
||
- Created the persistent Japanese manual design and Computer Use capture script at `docs/V38_BEGINNER_MANUAL_PLAN_AND_CAPTURE_SCRIPT_JA.md`. It is the canonical restart point for the later image-led V3.8 beginner guide.
|
||
- The plan reuses the current exact `README_JA.md` Review walkthrough, separates Quick Start from reverse-lookup help, defines stable Shot IDs, capture prerequisites, environment records, filenames, abnormal-state coverage, a bug-report template, and completion checks.
|
||
- Final screenshots remain deferred until the V3.8 public UI, standard workflow, and launch treatment of the active drift investigation are stable. A later instruction to start the manual must read the plan first and verify current labels before capturing `V38-M01` onward.
|
||
- No Computer Use capture, runtime synchronization, backend operation, GPU generation, publication, commit, or push was performed in this planning step.
|
||
|
||
## Issue13 Texture-only RGB/VAE-delta Feedback Clamp (2026-09-06, FAIL / closed)
|
||
|
||
- Static prototype/policy/source-order checks passed `13/13`, but the real fp16 H3 Video VAE feasibility gate failed before any Diffusion Sampling. GPU placement OOMed on the first 22f decode, including tiled retry: peak allocated `14,764,005,888` and reserved `16,280,190,976` bytes on the 16GB device.
|
||
- Corrected `--cpu-vae` fallback was then attempted. After about five minutes its owned PID used `54,954,274,816` bytes working set, private allocation reached `163,648,221,184` bytes with paging, available RAM fell to about `6,658MB`, and Windows displayed a Python memory application error. The owned process was interrupted; available RAM recovered to about `56,262MB` and GPU free memory to about `14.9GB`.
|
||
- Therefore the decoded-RGB `encode(corrected)-encode(decoded original)` path is operationally infeasible here and also conflicts with the protected V2 rule against Video/Audio VAE decode between H3 chunks. Close this concrete candidate; do not run its `6x5` efficacy A/B, switch to int8-only, lower resolution, CPU forcing, or threshold tuning automatically.
|
||
- Production/source runtime is unchanged. No ComfyUI backend or listener was started; owned PIDs `9440` and `4472` exited. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-texture-clamp-cpu-design-20260905\REAL_VAE_PREFLIGHT_RESULT_JA.md`. New GPU work still requires fresh explicit confirmation. A5/general A8b/Release HOLD remain unchanged.
|
||
|
||
## Issue13 Schedule-Matched Disposable Prefix screen (2026-09-05, closed)
|
||
|
||
- Completion review faf8917a-b718-48e6-b40c-b380cc68c17a / reply26bc256c-bf10-49e8-b131-0d01e2740844 received/read: experimentPASS / this22f screenFAIL, no mandatory correction, pair closed. Report-only review; no independent media inspection. No next GPU phase authorized by this result.
|
||
|
||
- User explicitly approved proposal execution and one frozen6x5 pair after integration gates. Experiment validity PASS / specified22f mitigation FAIL: edge drift39.8675→56.3675 (+41.39%); contrast/highlight guards failed, saturation/gradient/shadow passed. Flow mean0.510734→0.534240 (+4.60%); not full motion/semantic AV certification. No sweep or60s promotion.
|
||
- Same savedGroup1 on both sides,10new physical groups total. A801.79s/B779.90s, each720f/30s. A rawAV12/12 exact vs baseline. Actual Core noise/seed/SIGMAS/inputmask/shape5/5 equal; conditioning/Assembly equal; B40H3 mask observations valid; cleanAV exports10boundaries exact. CPU14+protected85 PASS, assets10/sourceRegistry104/isolatedPython83 unchanged,Public7. Fullpytest not rerun.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-schedule-prefix-20260905\RESULT_JA.md`. Production/runtime unchanged; only isolated experiment and source state/history docs edited. A5/generalA8b/Release HOLD, no commit/push. Owned8190 PIDs13280/26320 and monitors stopped, listener offline; user8188 idle/reachable/unmodified. PreGPU fixed-task review26665e59 read before A; completion review pending at report creation.
|
||
|
||
## Issue13 R3B consolidated report / direction review (2026-09-05)
|
||
|
||
- User requested consolidation of all completed R3B tests and direction review, not new GPU or Production changes. Ten30s runs/57physical groups/300s output; all completed, no accepted mitigation. Existing analysis equal_5s_windows and MP4 SHA were rechecked10/10. Do not substitute 3x10 chunk-center metrics for common5s windows.
|
||
- Consolidated report: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-consolidated-20260905\ISSUE13_R3B_CONSOLIDATED_REPORT_JA.md`. GPU_TEST_MATRIX.json retains exact values/provenance. HistoricalR2.6 remains a strong diagnostic positive, not a failed experiment; Video/Audio fixed-tail and motion confounds remain. R3B did not establish handoff-count-only causality or a motion-preserving fix.
|
||
- Fixed H3情報チェック review received/read: turn4ee9d332-985c-49b5-8f3c-1e3c5186d3d2, reply7326b44b-f2b4-4828-a1f6-d8011f5113c3. Supports investigation progress/Production mitigation none/initiating bias unresolved. Reviewer read report only, not local media independently; broad exclusion language is restricted to the tested settings.
|
||
- Proposed priorities only: reporter-exact differential, saved-data temporal/region analysis, then design one controlled same-prefix intervention if justified. No new intervention method approved or ready; temporal observations alone do not establish causality. No additional GPU, backend launch, Production fix or publication in this consolidation. A5/general A8b/Release HOLD remain. Source Registry104/104 unchanged; source edits only state/history documentation.
|
||
|
||
## Issue13 R3B non-Turbo Reference20 pair (2026-09-05)
|
||
|
||
- User explicitly authorized only two matched non-Turbo20step30s Reference OFF/ON cases. Ref2VA pruned INT8, seed131325,6x5,Manual576x864,Standard/Balanced22,AudioON,FirstImage; no LoRA node executed. Production/runtime and general A5/A8b/Release HOLD unchanged.
|
||
- Experiment PASS / this Reference mitigation screen FAIL. OFF1771.584s, ON1854.748s; each720f/30s AV,6physical groups. Laplacian delta25.1840→39.2554 (+55.87% larger, not a global quality score); saturation drift0.00238→0.01139 and highlight drift failed guards. Contrast drift improved58.05%; gradient/shadow nearly equal. No candidate accepted; one scene/seed does not reject native Reference generally or isolate Turbo causality versus prior8step.
|
||
- Actual Core noise/seed/SIGMAS/mask/target-shape6/6 match; both21SIGMAS unchanged and modelpatches0. Clean Video/Audio prefixes each10/10 exact. Same FirstImage keyframe and persistent native Reference payload in all6 ON groups; Assembly equal except runtime start. Sampled identity/scene retained; similar ON poses are a caution, not proof of freeze. Full semantic AV/motion unqualified.
|
||
- Evidence/report: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-ref2va-standard20-20260905\ISSUE13_R3B_STANDARD20_REFERENCE_GPU_REPORT_JA.md`. Registry104/104, livePublic7/7 before/after,10assets unchanged,AST18,observerCPU anddiff-check PASS; fullpytest not rerun. Only state/history documentation changes. Existing Development mismatch1/254 remains separate.
|
||
- Dedicated8190 launcher29056/backend29368 stopped19:14:23JST; PID/listener/HTTP,conhost32972 andmonitor exit verified. User8188 untouched. Monitoring370samples,minRAM0.995GiB,no safetyinterrupt. OuterPowerShell stale LASTEXITCODE check falsely returned nonzero after successful shutdown; independent exit checks confirmed success, no GPU rerun.
|
||
- Current R3B total10runs/57physical groups, no accepted mitigation. Fixed-task completion review a9db6c24-27cc-4b15-b0c6-da908eb1702f received/read: experimentPASS/screenFAIL, no mandatory rerun. Clarified that Reference changes the first generated image too; compare each conditioning trajectory's own start-to-end change, not identical initial videos. Reviewer did not independently view local media. Approved pair closed; no additional GPU, multiseed/long-form, Production promotion or publication under this two-case approval.
|
||
|
||
## Issue13 R3B isolated context-noise GPU screen (2026-09-05)
|
||
|
||
- User explicitly approved the proposed single-intervention Video context-noise experiment with "許可、許可。" after the Ref2VA review. This supersedes the earlier waiting-for-approval status only for this isolated A/B. Product A8b, A5 and Release remain HOLD. Production/runtime must stay unchanged.
|
||
- Work directory: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-noise-20260905`. CPU10/10 + existing69 protection tests, observer and5-boundary Shadow passed. Fixed-task review received; corrected Gaussian0.05 definition to a standard-deviation multiplier (not maximum amplitude), then ran only the approved pair.
|
||
- GPU experiment PASS / this0.05 tapered Video-noise screen FAIL. OFF876.398s and ON830.900s, each720f/30s. Laplacian delta39.8675→43.9942 (+10.35% worse), saturation delta+47.25% worse; contrast/highlight/shadow partly improved. Motion proxy0.510734→0.336667 (−34.08%) and repeated sampled poses are cautions, not proof of complete freeze. No candidate accepted or Production promotion.
|
||
- Fresh OFF raw Video/Audio matches prior baseline12/12. OFF/ON first raw AV exact, actual packed Core noise/SIGMAS/derived seeds6/6 exact, Assembly geometry and exogenous conditioning exact. Intended interventions5/5; newest slot delta0; clean Video/Audio prefixes each10/10 exact. Recursive Audio diverges fromGroup3 through joint generation but is never directly modified.
|
||
- Registry104/104 and livePublic7/7 passed;10asset files rehashed unchanged. Source changes are only AGENTS, PROJECT_STATE, WORKLOG and the AGENTS hash row in Development Manifest. Pre-existing Registry-manifest mismatch remains1/254; package/Production source unchanged. Full pytest not rerun. Dedicated8190 launcher1972/child4128 stopped16:26:44JST; PID/listener/HTTP and monitor exit verified, user8188 untouched.
|
||
- Report: `ISSUE13_R3B_NOISE_GPU_REPORT_JA.md`. Current exploration totals8runs/45physical groups, no accepted mitigation. This narrowly approved pair is complete; do not infer approval for a parameter sweep, another intervention, multiseed/long-form, or Production changes. Fixed H3情報チェック completion review received/read in turn4832f123-abeb-47de-9d97-56e25840a79c: experiment PASS / this condition screen FAIL; no mandatory correction or additional GPU. Reviewer did not independently view local media. Issue13 remains unresolved; A5/general A8b/Release HOLD remain.
|
||
|
||
## Issue #13 R3B Ref2VA paired GPU screen (2026-09-05)
|
||
|
||
- Initial-screen review was received from fixed H3情報チェック task `6a94b91e-af7c-83e8-be18-ddc504f690c4` (turn `1ca3ed8b-f859-43d7-84ee-7459c3a87ce9`) before this phase. It supported matched Ref2VA Reference OFF/ON and retained the initial conclusion: no accepted mitigation; Audio OFF is mixed.
|
||
- Ref2VA checkpoint + matching Ref2V8step LoRA, seed131325,6x5s,Manual576x864,Standard/Balanced22,AudioON: OFF and ON completed720f/30s AV in888.343s/918.375s. Reference connection was the only sampling-input delta. Laplacian increase was13.1606 OFF versus27.7282 ON (6.75%/13.20% from each starting value); native Reference did not meet the predeclared mitigation screen. Some secondary metrics improved, but no multi-seed/long-form promotion or Production fix is justified by this pair.
|
||
- Raw Video/Audio prefixes each10/10 exact; actual public-call SIGMAS identical and unchanged; First Image latent exact between cases; same Reference payload present in all6 ON groups. Twelve pre-sampling Refine conditioning signatures were recorded without hooks or mutation. CPU noise replay6/6 matches are reconstruction, not actual-noise telemetry. Sampled identity/scene remained, but semantic AV and full temporal naturalness remain unqualified.
|
||
- Report: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-ref2va-20260905\ISSUE13_R3B_REF2VA_GPU_REPORT_JA.md`. Six current-environment30s cases/33physical groups total across both screens; Issue13 remains unresolved. Pair completion review was received and read (turn `b916ed28-cb5f-4587-8456-4a7bff915049`): experiment PASS / mitigation screen FAIL; no candidate promotion, no additional mandatory GPU for this screen. Next requires reporter-exact workflow or explicit single-intervention experimental approval. Context-noise interventions remain subject to A8b HOLD; generic GPU authorization does not waive its release conditions. A5 and Release HOLD remain.
|
||
- Production/source runtime files unchanged; only project state/history recorded. Registry source/stage104/104 and live Public7/7 passed. Development Manifest's pre-existing stale REGISTRY_MANIFEST.sha256 entry (1/254 mismatch) remains separate. Full pytest not rerun. Dedicated8190 PIDs5792/3228, listener/HTTP and monitor exit verified at15:13:36JST, user8188 untouched. No commit/push/tag/release/publish.
|
||
|
||
## Issue #13 R3B initial GPU screen (2026-09-05)
|
||
|
||
- GPU execution completed for four matched current-environment30s cases:6x5 baseline,3x10,6x5 with native Reference Image1,6x5 with Audio Continuity OFF. All21 physical groups and four720f/30s AV outputs succeeded; no accepted Issue13 mitigation yet. Longer chunks and clean Reference did not meet the frozen screen. Audio OFF reduced Laplacian drift41.83% but increased saturation drift51.25%; keep it diagnostic, not a recommended default.
|
||
- Current FL2VA8step LoRA is not the unavailable historical4step LoRA. This is not reporter-exact R3. A/D first raw AV group is bit-exact; Video prefixes17/17 and Audio-ON prefixes12/12 are exact. CPU reconstructed noise parity is evidence replay, not actual Sampling-time telemetry. One seed/scene only; long-form, multiple seeds, semantic AV and wrapper matrix remain untested. A5/A8b remain HOLD.
|
||
- Evidence/report: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r3b-20260905\ISSUE13_R3B_INITIAL_GPU_SCREEN_REPORT_JA.md`. User explicitly authorized sharing this GPU audit only with pinned H3情報チェック task `6a94b91e-af7c-83e8-be18-ddc504f690c4`. Completion review was received before the paired Ref2VA-model native Reference gate; see the later phase above.
|
||
- Source/runtime Production was unchanged. Dedicated8190 PIDs36076/28936 and its monitor were stopped and checked offline at14:15JST; user8188 remained live and untouched. Registry104/104 passes. Development Manifest has one pre-existing stale entry for REGISTRY_MANIFEST.sha256 (same bytes as pre-GPU snapshot), retained for separate Release Track correction. Full pytest was not rerun in this GPU screen. Release HOLD and publication prohibition remain.
|
||
|
||
## V3.8 Review Fixed-Seed Guard (2026-09-05)
|
||
|
||
- `Run = Review Each Chunk` now requires the visible `Control After Generate = fixed` setting. With `randomize`, `increment`, or `decrement`, the Main card shows `Set Control After Generate to fixed`, explains that Base Seed must remain unchanged, and the frontend cancels Queue with the same exact instruction instead of silently creating another unrelated Chunk 1 revision.
|
||
- `Try this chunk again` continues to create a distinct Take through the existing Review variation/nonce contract; users must not randomize Base Seed for it. Full Run and V3.7 or earlier nodes retain their existing Control After Generate behavior.
|
||
- Run Storage reload now runs after `execution_success`, `execution_error`, and `execution_interrupted`, with the existing node-local execution callback retained as fallback. This makes completed, failed, and interrupted persistence states visible without changing Queue or Review intent.
|
||
- The immediate README walkthrough now begins with `Control After Generate = fixed`. A separate image-led, exact-label, one-action-per-step beginner guide remains a required follow-up before release; it must describe a known-good `2 × 5s` path without internal terminology or inferred UI labels.
|
||
- Focused Review/UI regression passed `116/116`; full pytest passed `1127/1127`; JavaScript syntax, Development Manifest `254/254`, Registry Manifest `104/104`, runtime verifier/public surface `7/7`, source/runtime SHA `2/2`, all four live served-extension markers, and `git diff --check` passed.
|
||
- Public/backend schema, widget/socket order, Sampling v5, Run Storage v3, Branch/Take provenance, State/Session, Assembly, and GPU behavior remain unchanged. Release / Registry readiness remains **HOLD** pending package staging after this UI/documentation change.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-review-seed-guard-filtered-20260905_024733-0a7c670` (`11/11`, mismatch `0`).
|
||
|
||
## V3.8 Review Completion Refresh + Step-by-step Guide (2026-09-05)
|
||
|
||
- Confirmed the reported `2 × 5s` run was configured correctly and that the first 5-second output is the intentional first Review unit. Run Storage contained a valid canonical `review_ready` Chunk 1 with 33 Takes; the defect was the stale frontend view after execution, not missing backend generation or storage.
|
||
- The V3.8 frontend now reloads Run Storage on ComfyUI's global `execution_success` event, with the existing per-node `onExecuted` hook retained as a fallback. The first load is immediate and one bounded 250 ms follow-up covers filesystem/UI timing without changing a widget, choosing an action, or automatically starting Queue.
|
||
- `README.md` and `README_JA.md` now give a literal `Chunks = 2`, `Seconds per Chunk = 5`, `Total Length = 10 seconds`, `Run = Review Each Chunk`, `Progress = On — Resume and Takes available` walkthrough. It explains the intentional 5-second first result, all three visible Review buttons, the required second Queue, Back to Settings/Return to Review, and common mistakes using the exact UI labels.
|
||
- Focused Review/UI regression passed `86/86`; full pytest passed `1126/1126`; JavaScript syntax, Development Manifest `254/254`, Registry Manifest `104/104`, runtime verifier/public surface `7/7`, source/runtime SHA `2/2`, live served-extension markers, and `git diff --check` passed.
|
||
- Public/backend schema, serialized widgets, Review intent, Run Storage v3, Branch/Take provenance, Sampling v5, State/Session, Assembly, and GPU behavior are unchanged. Release / Registry readiness remains **HOLD** pending package staging after this frontend/documentation change.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-review-refresh-readme-filtered-20260905_022621-0a7c670` (`10/10`, mismatch `0`).
|
||
|
||
## V3.8 Single-chunk Review Guidance (2026-09-05)
|
||
|
||
- `Run = Review Each Chunk` with `Chunks = 1` no longer presents a misleading `Ready to Queue` card. The Main card now explains that the first chunk is also the last and directs the user to set `Chunks` to `2` or more before Queue to use Review and Continue.
|
||
- The `Chunks` and `Run` mouse-over help carries the same instruction. This is diagnostic guidance only: values are not changed automatically and Queue is not blocked.
|
||
- Public/backend schema, serialization, Run Storage v3, Review/Branch semantics, Sampling v5, State/Session, Assembly, and GPU behavior are unchanged. Release / Registry readiness remains **HOLD** pending package staging after the UI change.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-review-single-chunk-guidance-filtered-20260905_010500-0a7c670` (`10` exact files, mismatch `0`).
|
||
|
||
## V3.8 Review Back to Settings (2026-09-05)
|
||
|
||
- A canonical `review_ready` screen now includes `Back to Settings`. It restores the normal Main controls without changing the pending Review action, selected Take, canonical branch, Queue, or any serialized backend value. `Return to Review` returns to the three Review actions.
|
||
- The normal `Ready to Queue` card shows the still-pending chunk Review while Settings are open. A newly completed review revision opens the Review screen automatically; the view toggle itself is transient and non-serialized.
|
||
- This is Frontend presentation only. Public schema/sockets, Sampling v5, Run Storage v3, State/Session, Assembly, Branch/Take provenance, Seed/SIGMAS, and GPU behavior are unchanged. Release / Registry readiness remains **HOLD** pending package staging after the UI change.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-review-back-to-settings-20260905_002018-0a7c670` (`8` exact files, mismatch `0`).
|
||
|
||
## V3.8 Main Core Controls + Card Alignment (2026-09-04)
|
||
|
||
- Restored `Prompt Format`, `Continuity`, `Base Seed`, `Control After Generate`, and `Audio Continuity` as the first five normal Main controls, using readable facade labels while preserving the existing serialized backend widgets and values. These controls no longer require opening Advanced.
|
||
- Fixed the custom Ready/warning/Review card draw contract. The card now always draws with its declared 62/68 px height instead of accepting an incorrect 20 px widget-height argument from the canvas callback, so the rounded border no longer crosses the headline or detail text.
|
||
- Focused UI/schema passed `86/86`; full pytest passed `1124/1124`; JavaScript passed `7/7`; Development Manifest passed `254/254`; Registry Manifest passed `104/104`; source/runtime UI SHA matched; runtime verifier/public surface passed `7/7`; and `git diff --check` passed. Public Schema, sockets, Sampling, Run Storage, State/Session, Assembly, and GPU behavior remain unchanged.
|
||
- Pre-change snapshots: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-main-core-controls-card-alignment-source-filtered-20260904_223526-0a7c670` (`10` exact files, mismatch `0`) and `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-main-core-controls-card-alignment-runtime-w-filtered-20260904_223526` (`2` exact files, mismatch `0`). The standard helper stopped only on the known `.pytest_cache` ACL.
|
||
|
||
## V3.8 Explicit Main Controls (2026-09-04)
|
||
|
||
- Replaced the over-automatic Main facade with explicit, independent `Chunks` and `Seconds per Chunk` controls. `Total Length` is now a read-only calculation only; changing it cannot recalculate either generation setting. Values such as `3 chunks × 10 seconds` and `2 chunks × 15 seconds` are directly configurable again.
|
||
- Removed the permanent Connected Inputs badges and automatic Input Mode row. `Size Source` is always visible and remains exactly `First Image` or `Manual` until the user changes it. First Image uses the existing Resolution choices; Manual shows the exact Width and Height directly. The facade no longer converts an inactive First Image into an apparently user-selected custom frame size.
|
||
- If `Size Source = First Image` has no active First Image because it is missing, disabled, or bypassed, the Queue summary reports `Check First Image` and the exact fallback size without silently changing the selection. Reference Image and Video Guide size controls remain conditional inside Advanced.
|
||
- Review-ready behavior remains the approved amber state with three explicit Queue actions. Public Python schema, sockets/order, serialized widget values/order, Sampling v5, Run Storage v3, State/Session, Assembly, Branch/Take, Seed/SIGMAS, and GPU behavior are unchanged. Focused UI/schema passed `86/86`; full pytest passed `1124/1124`.
|
||
- The standard pre-change snapshot stopped only on the established `.pytest_cache` ACL. Accepted filtered snapshots are `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-explicit-controls-source-filtered-20260904_214419-0a7c670` (`10` exact files, mismatch `0`) and `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-explicit-controls-runtime-w-filtered-20260904_214419` (`2` exact files, mismatch `0`). Release / Registry readiness remains **HOLD** pending package staging after this UI change.
|
||
|
||
## V3.8 Mock-faithful Main UI (2026-09-04, superseded by Explicit Main Controls)
|
||
|
||
- Reworked the presentation-only V3.8 Main facade to follow the approved mock hierarchy: active-input badges, automatic `Input Mode`, one `Video Length` selector plus `chunks × seconds` detail, conditional Output/Quality or Frame Size, Run, Progress, a Queue-ready summary card, and one Advanced disclosure. Existing serialized widgets remain behind that facade and keep their original values/order.
|
||
- Input detection now follows effective graph state. A connected First Image, Video Guide, or Audio loader in ComfyUI bypass/never mode (`2`/`4`) or with its enable widget OFF is shown as inactive. This prevents the public workflow's bypassed optional loaders from falsely changing T2VA to I2VA or exposing irrelevant controls.
|
||
- Review mode uses the second approved mock state only when Run Storage reports the canonical revision as `review_ready`: an amber chunk-ready card, `Use it and continue`, `Try this chunk again`, and `Use it and finish the rest`, followed by Render History and Advanced. Selecting an action does not auto-Queue and does not change the canonical branch until the normal Queue runs.
|
||
- Public Python schema, input sockets/order, serialized widget values/order, Sampling v5, Run Storage v3, State/Session, Assembly, Branch/Take, Seed/SIGMAS, and GPU behavior are unchanged. Focused UI/schema passed `86/86`; full pytest passed `1124/1124`. Live visual acceptance remains for the user's next ComfyUI reload because the isolated browser route could not be used without joining an existing Edge process; no user tab was touched.
|
||
- Pre-change snapshots: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-mock-faithful-ui-source-20260904_201500-0a7c670` (`327` files, mismatch `0`) and `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-mock-faithful-ui-runtime-w-20260904_201500` (`105` files, mismatch `0`). The dedicated validation backend on port `8190` and its exact temporary Edge profile were stopped; ports `8190`, `8188`, and `9223` were not listening at cleanup. Release / Registry readiness remains **HOLD** pending package staging after this UI change.
|
||
|
||
## V3.8 Complete Mouse-over Help (2026-09-04)
|
||
|
||
- Added non-empty mouse-over help to all 12 Main facade controls, all nine visible Advanced controls, Review/Render History actions, and all 47 required/optional public V3.8 Sampler inputs. Tooltip text is stored on both the widget and widget options so current ComfyUI frontend paths can display it consistently.
|
||
- `Output Size` now explains the selected behavior directly: First Image preserves the connected image aspect for I2VA/FL2VA, while Manual exposes exact 32-aligned Width/Height for T2VA, missing-First-Image workflows, or explicit canvases. Resolution, Resume & History, Reference Image Size, and Video Guide Size help also updates with the current selection.
|
||
- This is help text and presentation metadata only. Widget names/order/values, sockets/types, Public `7/7`, Sampling v5, Run Storage v3, State/Session, Assembly, Branch/Take, Seed/SIGMAS, and GPU behavior remain unchanged. Focused UI/schema passed `85`; full pytest passed `1124/1124`; JavaScript passed `7/7`; Development Manifest `254/254`; Registry Manifest `104/104`.
|
||
- Pre-change snapshots: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-all-hover-help-source-filtered-20260904_192259` (`326` files, mismatch `0`) and `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-all-hover-help-runtime-w-filtered-20260904_192318` (`105` files, mismatch `0`). The four changed Registry files are synchronized to ComfyUI_W with source/runtime SHA `4/4`; the live extension route returns the new help code. The already-running backend was deliberately not restarted, so new Python socket help becomes visible after the user's next restart. Eight unrelated source/runtime Registry mismatches predated this task and were preserved for a separate full-runtime sync decision. Release / Registry readiness remains **HOLD** until a tooltip-inclusive package staging Gate passes.
|
||
|
||
## V3.8 Intuitive Main UI Refinement (2026-09-04)
|
||
|
||
- The compact facade now exposes `Generation = Full Video | Review by Chunk`, integer `Chunks`, compact `Seconds per Chunk`, `Output Size`, `Resolution`, `Resume & History = Off | On`, connection-specific Reference/Video Guide sizing, and one `Show/Hide Advanced Settings` button. Seed remains available inside Advanced Settings.
|
||
- Chunks and Width/Height display without decimal places. Seconds display only a meaningful single decimal, so `3.000` and `5.000` render as `3` and `5` while values such as `5.5` remain available. `Review by Chunk` automatically enables the unchanged backend `Save + Auto Resume` requirement.
|
||
- Review controls no longer appear merely because Review mode was selected. Before the first Queue, the UI asks the user to create the first chunk. Only a canonical Run Storage revision with backend status `review_ready` reveals `Keep & Continue`, `Try This Chunk Again`, and `Finish Automatically`; selecting one leaves a persistent Queue instruction instead of only a momentary button flash.
|
||
- Empty Render History is hidden. Once at least one backend Take exists, the collapsed History button becomes visible and retains the existing N2c Take selection/actions. Video Guide and Reference Image sizing remain connection-driven.
|
||
- This is Frontend presentation only. The serialized V3.8 widget order and values, Python/public schema, seven public nodes, Sampling v5, Run Storage v3, Branch Provenance, State/Session, Assembly, Seed/SIGMAS, and GPU behavior are unchanged. Focused UI passed `15`; full pytest passed `1120/1120`.
|
||
- Source/runtime snapshots: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-intuitive-ui-prototype-source-filtered-20260904_144147` (`915/915`, mismatch zero) and `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-intuitive-ui-prototype-runtime-w-20260904_144128` (`179` files). The user-provided rollback ZIP remains unchanged at `D:\StabilityMatrix\Data\Packages\ComfyUI_W\custom_nodes\ComfyUI-H3-Continuum38_924.zip`, SHA-256 `6FF68ECA71DA9459C15EAB5B31480D65F03E2734E3ED3A2D86D7DCB024CB802D`.
|
||
- Refinement snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-intuitive-ui-refinement-20260904_153802` (`9` exact source/runtime files). Synced only `web/project_id.js` and `REGISTRY_MANIFEST.sha256` to ComfyUI_W; hashes match, the live extension route returns HTTP 200 with every new marker, and runtime verification passes public `7/7`. User visual acceptance, README/CBT images, and package staging remain later work. **Release / Registry readiness = HOLD.**
|
||
|
||
## V3.8 Resolution / Size Source UX Refactor PASS (2026-09-03)
|
||
|
||
- The user moved V3.8 test videos and test artifacts to `D:\output\video\comfy_video\v38`; the verified directory currently contains 868 files. Treat it as the current V3.8 output archive and preserve its contents.
|
||
- Main V3.8 now presents only `Size Source = First Image | Manual`. First Image preserves the connected First Image aspect and reuses the existing Draft 0.30 MP, Balanced 0.60 MP, Native 768, Custom MP, 32-pixel alignment, and 1344 long-edge-cap resolver. Manual accepts exact aligned Width and Height and does not apply the preset.
|
||
- Width and Height are appended after every existing serialized V3.8 widget. They remain visible, are read-only in First Image mode, and become editable in Manual mode. The frontend updates them from exact available First Image preview geometry before Queue; the backend remains authoritative and reports the final resolved canvas in Status.
|
||
- The prior `aspect` key and its old values remain accepted as a hidden compatibility field. Frontend migration maps Auto to First Image and maps Landscape, Portrait, and Square to Manual using the exact former resolved canvas. API prompts that omit the new fields retain the legacy resolver path.
|
||
- If First Image is selected but no usable First Image reaches the sampler, execution follows the project fail-open rule: the displayed Manual Width and Height are used deterministically and Status states the fallback. Reference Image and Video Guide are never implicit size sources. FL2VA without First Image therefore uses the same explicit Manual fallback.
|
||
- Public node count remains `7/7`. The only Python schema change is the approved append-only `size_source`, `width`, and `height` UI contract; existing keys/order/sockets remain intact. Sampling Contract v5, Seed/SIGMAS, physical grouping, Run Storage v3, Branch Provenance, State/Session, Assembly, Driving Audio, Reference Audio, Second Pass/O1, and Terminal Merge are unchanged.
|
||
- Validation passed focused `107`, full pytest `1117`, JavaScript `7/7`, Development Manifest `253/253`, Registry Manifest `102/102`, public workflow `19` nodes / `25` links / orphan `0`, protected contracts `16/16`, Python compile, and `git diff --check`.
|
||
- Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-resolution-size-source-ux-refactor-filtered-20260903_152559-0a7c670` (`320/320`, mismatch zero). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-resolution-size-source-ux-20260903_153907`.
|
||
- Accepted snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\accepted-v38-resolution-size-source-ux-final-20260903_160000-0a7c670` (`320/320`, mismatch zero). Formal result: **Resolution / Size Source UX Refactor = PASS**. Because the last staged ZIP predates this UX/schema/workflow change, **Release / Registry readiness returns to HOLD until a new package staging Gate passes**. GPU was unnecessary. On user request, ComfyUI_W was preserved in exact 199-file runtime snapshot `pre-v38-resolution-size-source-runtime-w-sync-20260903_163000`, then synchronized from the current 102-entry Registry Manifest. The initial four-file-only attempt exposed the expected missing N2b dependency `branch_provenance.py`; the complete manifest sync resolved it. Restarted only the command-line-verified ComfyUI_W backend at `127.0.0.6:8188`; live `object_info` and frontend routes now pass public registration `7/7`, Size Source/Width/Height schema, and current `project_id.js`. No commit, push, tag, release, or publish was performed.
|
||
|
||
## ComfyUI_W Reference-Audio / Installer Sync PASS (2026-09-03)
|
||
|
||
- Treated the already-correct ComfyUI_W implementation as authoritative for this limited synchronization. `reference.py` now adds each bundled Reference Audio individually and in order when Reference Image conditioning is also present; legacy single-Audio behavior remains unchanged.
|
||
- Added the public-surface verifier helper to the package allowlist, made the Windows installer stop explicitly when that helper is absent, and included the runtime dependency `v3/reference_video.py`. Registry Manifest is now `104/104`; Development Manifest remains `253/253`, both with zero mismatch.
|
||
- Focused source regression passed `18`; it includes the new Reference Image + Audio 1/2/3 presentation and `minimax_refs` ordering test. Stage-only import and runtime verifier both passed with public `7/7`. No runtime synchronization, backend/frontend operation, GPU, commit, push, tag, release, or publish occurred. Release / Registry readiness remains HOLD pending official package staging.
|
||
|
||
## V3.8 Final Package Staging Re-Gate PASS / Release readiness GO (2026-09-03)
|
||
|
||
- The latest 320-file working tree was staged through official `comfy-cli 1.18.0`. The first current-tree archive exposed one packaging-only gap: historical validation/design documents, old-version UI images, `.github`, `.gitignore`, and `pytest.ini` were still included.
|
||
- The minimal correction changed only `.comfyignore` plus the two Manifests. Historical/internal docs and repository-only configuration are excluded; only the current N2b, O1, and V3.8 release/migration contracts are retained. Production Python/JavaScript/workflow/schema/Sampling/Run Storage/State/Session/Assembly code is unchanged.
|
||
- Official `comfy node validate` and `pack` exited zero. The final archive contains 103 files, is 349704 bytes, and has SHA-256 `9675373D8BD37A345D4467BF3BDB6E869EA3BFE051B730E316AE338A60B0D97F`.
|
||
- Source→stage and stage→archive missing/extra/mismatch are all zero. Registry Manifest is `102/102`; Development Manifest is `253/253`. Duplicate, unsafe, zero-byte, banned internal, secret, user-specific-path, and temporary-log findings are all zero.
|
||
- Isolated stage-only import passed with no repository fallback. Public class/display/category/schema is `7/7`; AUDIO-R1 helper/backend/frontend/socket and O1 Second Pass finite validation/contract are present. The package contains exactly the two required `tools/` runtime files and one V3.8 public workflow with no orphan/duplicate links or old/experimental nodes.
|
||
- Validation passed focused `126`, full pytest `1086`, JavaScript `7/7`, runtime verifier, protected contracts `16/16`, and `git diff --check`. GPU was not run.
|
||
- Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-final-package-staging-regate-filtered-20260903_141442-0a7c670` (`320/320`, mismatch zero). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-final-package-staging-regate-20260903_141519`.
|
||
- The designated `H3情報チェック` task accepted **V3.8 Final Package Staging Re-Gate = PASS**, found no mandatory correction or additional Gate, and approved **Release / Registry readiness = GO**. Commit, push, tag, GitHub Release, Registry publish, and all other publication remain separate explicit-approval actions and were not performed.
|
||
|
||
## Issue #19 Chunk Boundary Continuity PARTIAL REPRODUCTION (2026-09-03)
|
||
|
||
- The reported severe video/audio cut resembling separately joined clips was not reproduced under the wrapper-free V3.8 baseline. Dedicated `8190` completed `1x5`, `2x5`, `3x5`, `2x10`, and matched-seed `2x5` explicit-prompt cases; optional `3x10` was safely omitted after the `2x10` run reduced available system RAM to about `0.004 GiB` without exposing a new anomaly.
|
||
- Video and Audio latent prefixes were bit-exact at all five physical boundaries. Separately decoding each physical group nevertheless made the same 22-frame RGB overlap non-bit-exact at `5/5` boundaries and increased the visible-join frame MAE by `10.657–21.899%` relative to the raw same-group prefix-to-generated transition. Finalized boundaries remained non-outliers (`0/5`) with SSIM `0.953167–0.972468`; no severe visual cut or trajectory restart was observed.
|
||
- Audio had no silence or dropout. Five-second Seam-OFF splice jumps remained below each case's ordinary adjacent-sample P99.9, and existing Audio Seam Auto reduced them from about `0.0958–0.0964` to `0.00354–0.01428`. The accepted FL2VA Terminal `[1]+[2,3]` evidence also showed no new video or audio cut and preserved Terminal atomicity.
|
||
- Issue #13 remains a separate cumulative Group 4–6 recursive-feedback pattern. The immediate weak decode-context effect observed here starts at Group 1→2 and did not worsen through three groups. Prefix corruption is rejected for both; Fixed Prefix remains diagnostic-only and was not run for Issue #19.
|
||
- Formal cause status: **separate-decode context asymmetry is a reproducible low-level candidate, but causal contribution to the reporter's severity is unproven**. Production remains unchanged. Further promotion requires either reporter-exact reproduction or a cache-only A/B of the same saved latent through separate versus consistent-context/overlap-aware Video VAE decode.
|
||
- Validation passed focused `72`, full pytest `1086`, development Manifest `253/253`, Registry Manifest `141/141`, runtime/public export `7/7`, pre-Gate snapshot comparison `320/320`, and `git diff --check`. Dedicated `8190` was stopped after Queue `0/0`; the temporary probe was removed; `8188` and the user Frontend were untouched.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue19-chunk-boundary-continuity-20260903_130500`. Pre-Gate snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-issue19-chunk-boundary-reproduction-filtered-20260903_130348-0a7c670` (`320/320`, mismatch zero).
|
||
- The designated `H3情報チェック` task formally accepted **ISSUE19 PARTIAL REPRODUCTION**, found no mandatory correction or additional GPU requirement for this Gate, and agreed that no Production change is justified. Release / Registry readiness remains HOLD; no commit, push, tag, or release was performed.
|
||
|
||
## V3.8 AUDIO-R1 Multi Reference Audio 1/2/3 PASS (2026-09-03)
|
||
|
||
- Core ComfyUI 0.34.2 formally supports up to three standalone `ref_audios`, presents them to the tokenizer in connection order as `<Audio 1>`, `<Audio 2>`, and `<Audio 3>`, and packs every independently encoded `[1,32,2,T]` latent into matching ordered `ref_audio` rows. No Continuum-specific conditioning format was added.
|
||
- The public surface is now exactly seven nodes. `H3 Continuum Reference Audios` owns three optional ordered AUDIO inputs plus one shared Core Audio VAE and emits one `H3_CONTINUUM_AUDIO_REFERENCES` bundle. V3.8 appends only `Audio References (Optional)` after every existing Sampler socket; V3.7 and earlier schemas remain unchanged.
|
||
- The helper requires gap-free connection order so UI numbering and Core prompt numbering cannot diverge. Each source is validated, resampled, hashed, cached, and VAE-encoded independently. Tokenizer items and `minimax_refs` preserve the same order. Audio 1/2 swapping changes the ordered bundle contract and Run Storage identity.
|
||
- The legacy single `Reference Audio (Optional)` plus `Reference Audio VAE (Optional)` path remains available and retains its v1 source contract, identity hash, conditioning payload, and saved-workflow socket order. Helper and legacy audio cannot be connected together. An unused legacy VAE remains ignored.
|
||
- Run Storage Schema v3 is unchanged. Its existing Reference Audio contract field carries the bundle contract only for helper runs; the helper upstream route fingerprints its VAE and ordered sources. The legacy upstream graph descriptor and SHA are bit-exact against the pre-AUDIO-R1 snapshot. Sampling Contract v5, State/Session, Branch Provenance, Assembly Plan, Driving Audio, Second Pass/O1, Seed/SIGMAS, physical grouping, and Terminal Merge are unchanged.
|
||
- Validation passed focused `70`, full pytest `1086`, development Manifest `253/253`, Registry Manifest `141/141`, runtime/public export `7/7`, native three-audio PackedLayout `18` reference rows, Python compile, protected contracts `16/16`, and `git diff --check`. GPU was unnecessary because Core-native layout/conditioning is exercised directly and Sampling tensors, seeds, grouping, and assembly were not changed.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-audio-r1-multi-reference-audio-filtered-20260903_113210-0a7c670` (`319/319`, mismatch zero). Accepted snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\accepted-v38-audio-r1-multi-reference-audio-filtered-20260903_115310-0a7c670` (`320/320`, mismatch zero). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-audio-r1-multi-reference-audio-20260903_114811`.
|
||
- Formal result: **AUDIO-R1 PASS / V3.8 Multi Reference Audio complete**. Release / Registry readiness remains HOLD until an AUDIO-R1-inclusive package staging dry-run passes. No runtime synchronization, live backend/frontend operation, GPU Sampling, V3.7 backport, compatibility reproduction, chunk-boundary investigation, Issue #13/A5/A8b work, commit, push, tag, or release was performed.
|
||
|
||
## V3.8 O1 Open Integration Contract PASS (2026-09-03)
|
||
|
||
- The supported external processing route is now explicit: `Sampler V3.8 -> optional external Video LATENT processor/upscaler -> Second Pass -> Core/third-party Video+Audio Decode -> Finalize -> Core Save Video`. No All-in-One node or new public surface was added.
|
||
- `H3 Continuum Second Pass` is the single supported latent-processing boundary. External processing may preserve or enlarge Video spatial geometry only; it must preserve physical-group count/order, atomic Terminal Merge groups, Video B/C/T, one common target H/W, matching first-pass Audio, and the original Assembly Plan.
|
||
- Second Pass now rejects NaN/Inf in incoming Video or Audio LATENT tensors before Sampling. Existing group count/order, rank, 24-channel, batch, temporal, spatial, common-H/W, and Audio-shape validation remains intact. Valid-input Sampling arguments, RefineSchedule, Seed derivation, conditioning, and output schemas are unchanged.
|
||
- First-pass Audio LATENT objects remain exact object passthrough and temporary refined Audio remains discarded. Geometry updates use a copied Assembly Plan and retain all non-target-geometry fields; Second Pass does not import or mutate Run Storage.
|
||
- Finalize remains decoder-neutral: it consumes only public ComfyUI `IMAGE` and `AUDIO` values plus the Assembly Plan. Core VAE, TensorRT-backed VAE, other compatible decoders, encode/decode tiling, fallback, and device lifecycle remain external responsibilities.
|
||
- SageAttention, Sol-Attn, Spectrum, and comparable accelerators remain external MODEL wrappers. Continuum uses public ModelPatcher APIs, imports no accelerator-private modules, does not select attention backends, and performs no global accelerator monkey patch.
|
||
- Formal contract: `docs/V38_OPEN_INTEGRATION_CONTRACT.md`. Regression covers valid external spatial processing, invalid group count/rank/temporal/downscale/Audio shape/NaN/Inf, Terminal Merge atomicity, Audio object identity, Assembly Plan immutability, public Second Pass/Finalize schemas, and absence of optional accelerator or Run Storage imports.
|
||
- Validation passed focused `80`, full pytest `1075`, public/release regression `20`, JavaScript `7/7`, development Manifest `252/252`, Registry Manifest `141/141`, source runtime verifier/public export `6/6`, protected contracts `16/16`, Python compile, and `git diff --check`. GPU was unnecessary because valid-input Second Pass Sampling and processing behavior did not change.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-o1-open-integration-contract-filtered-20260903_032012-0a7c670` (`317/317`, mismatch zero). Accepted snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\accepted-v38-o1-open-integration-filtered-20260903_033231-0a7c670`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-o1-open-integration-20260903_033231`.
|
||
- Formal result: **O1 PASS / Open Integration boundary established**. N1, N2b, and N2c remain PASS. Release / Registry readiness remains HOLD until an O1-inclusive package staging dry-run passes. Runtime synchronization, backend/frontend operation, GPU Sampling, Audio multi-reference, Issue #13, A5, A8b, release work, commit, push, tag, and release were not performed.
|
||
|
||
## V3.8 N2c Render History / Take UX Completion PASS (2026-09-03)
|
||
|
||
- Production view now presents the complete stored physical-group Take catalog in a read-only multiline `Render History / Takes` surface. Every Take shows its per-group Take number/count, short immutable revision ID, variation nonce, parent revision, Selected marker, Canonical marker, and canonical Head marker. Terminal Merge groups are labeled as one atomic range.
|
||
- Selection defaults to the backend canonical head and Previous/Next walks the complete stored catalog without modifying `active_revisions`. `Use This Take` and `Continue From Here` only prepare the existing one-shot Run Storage v3 action and still require normal Queue; selecting alone never changes the canonical branch and no automatic Queue path was added.
|
||
- The Continue summary follows the selected revision's backend-provided parent chain and reports the exact reusable physical-group prefix plus later stored groups that will regenerate. It does not infer accepted/generated state from frontend widgets. Run Storage OFF is explicitly read-only until an action enables storage.
|
||
- History reloads from `project.json` on Production workflow/view load, Run Storage identity changes, and node execution completion, covering Continue, Regenerate Current, Use This Take, and backend restart. Missing history is a clean empty state. Schema v2 project data is shown as legacy/read-only until the existing backend non-destructively synthesizes Branch Provenance on a normal Queue; stale Take selectors are cleared safely.
|
||
- All N2c helpers remain transient and Production-only. Basic hides them and workflow serialization excludes them. Public Python Schema/widget order/sockets, Run Storage Schema v3, Branch Provenance v1, Sampling Contract v5, State/Session, Assembly, Seed/SIGMAS, and the six-node public surface are unchanged.
|
||
- Validation passed focused `110`, full pytest `1064`, JavaScript behavior/syntax, development Manifest `250/250`, Registry Manifest `140/140`, source runtime verifier/public export `6/6`, and protected-contract parity; `git diff --check` exits zero. GPU was unnecessary because only the frontend presentation of already validated Run Storage v3 facts changed.
|
||
- Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-n2c-render-history-take-ux-filtered-20260903_025506-0a7c670` (`317/317`, SHA mismatch zero). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2c-render-history-take-ux-20260903_030627`.
|
||
- Formal result: **N2c PASS / Render History & Take UX complete at the approved scope**. N1 and N2b remain PASS; N2 remains unblocked. Release / Registry readiness remains HOLD until an N2c-inclusive package staging dry-run passes. Open Integration, Audio multi-reference, Issue #13, A5, A8b, P1a/P2, commit, push, tag, and release were not entered.
|
||
|
||
## V3.8 N2b Branch Provenance Contract v1 PASS (2026-09-03)
|
||
|
||
- Run Storage Schema v3 now persists immutable physical-group revisions with exact parent revision, seed/nonce identity, source storage revision, branch cut, active revision map, and canonical head. Normal groups remain one chunk; Terminal Merge remains one atomic physical group.
|
||
- `Use This Take` changes only the canonical selection and performs no Sampling. `Continue From Here` preserves the selected ancestral prefix exactly and opens a new nonce boundary at the next physical group. Older sibling branches remain addressable and are never overwritten.
|
||
- Schema v2 runs remain readable through a synthesized in-memory legacy provenance chain and are not rewritten on read. New Schema v3 storage identities are separate from the unchanged Sampling Contract v5 identity, preventing v2 manifest overwrite or false chunk reuse.
|
||
- Canonical updates occur only after a finalized `review_ready` or `complete` manifest is valid. In-progress/interrupted revisions are excluded from Take candidates; missing/corrupt `project.json` recovers deterministically from validated finalized manifests. There is no automatic history deletion or GC.
|
||
- Production view adds transient Render History / Takes controls for Previous, Next, Use This Take, and Continue From Here. Basic view hides them; workflow serialization removes the transient helpers. Actions still require the normal Queue operation and do not create a second execution path.
|
||
- State/Session schemas, Assembly Plan, Sampling Contract (`5`), seed/SIGMAS derivation, physical grouping, and public six-node registration are unchanged. GPU Sampling was not required because every generated region still delegates to the existing Sampling path; this Gate validates provenance, persistence, selection, restart, and frontend mapping on CPU/static paths.
|
||
- Validation passed focused `101`, full pytest `1062`, public-surface follow-up `7`, JavaScript `7/7`, development Manifest `250/250`, Registry Manifest `140/140`, source runtime verifier/public export `6/6`, and protected contract parity `9/9`; `git diff --check` exits zero. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2b-branch-provenance-20260903_024344`.
|
||
- Source snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-n2b-branch-provenance-contract-v1-filtered-20260903_020411-0a7c670` (`314/314`, SHA mismatch zero). No runtime synchronization, backend/frontend operation, commit, push, tag, release, GitHub, or Civitai action was performed.
|
||
- Formal result: **N2b PASS / N2 implementation boundary closed**. N1 remains PASS/Close. Audio multi-reference remains Civitai-reply HOLD; Issue #13, A5, and A8b remain separate; P1a/P2 remain Deferred. The prior staged archive predates N2b, so **Release / Registry readiness is HOLD until an N2b-inclusive package staging dry-run passes**; that separate Release Track was not run here.
|
||
|
||
## V3.8 N2 Branch / Take / Render History Audit HOLD (2026-09-03)
|
||
|
||
- Read-only audit found that existing Run Storage already preserves Smart Regenerate results as separate revision directories. `project.json` exposes revision/status/time/nonce/boundary/review-unit metadata, and each manifest retains per-chunk records plus seed/global transport contract facts. Terminal Merge review units remain atomic and disk revisions survive restart.
|
||
- GO-A/GO-B are not satisfied. Execution always chooses the latest valid compatible head by effective nonce, timestamp, and revision ID; no existing widget carries an exact revision ID. More importantly, the sampling identity supports only one reroll boundary and one nonce applied to every chunk at/after that boundary.
|
||
- This single-boundary contract cannot preserve `Chunk 3 Take B (boundary 3 / nonce 1)` while creating a new downstream branch at Chunk 4 with another nonce: boundary 4 makes Chunk 3 return to nonce 0, while boundary 3 / nonce 1 also keeps the old downstream lineage. Selecting a complete older head likewise reports sequence complete instead of invalidating only its later groups.
|
||
- A truthful read-only Take list could be frontend-only, but `Use This Take` and `Continue From Here` would require per-group provenance/parent revision plus branch-cut state, or another persisted exact active-revision selector. Either changes Run Storage format/identity validation and potentially Sampling Contract; both are explicit N2 stop conditions. No partial/cosmetic UI was implemented.
|
||
- Focused Review/Run Storage/Session tests passed `109`; development Manifest `247/247`, Registry Manifest `138/138`, pre/post source `314/314` unchanged before this state record, and `git diff --check` passed. Snapshot: `pre-v38-n2-branch-take-render-history-filtered-20260903_014541-0a7c670`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-n2-branch-take-audit-20260903_015018`.
|
||
- Formal result: **N2 HOLD / no implementation**. N1 remains PASS/Close. Production code, frontend/runtime, public Schema, State/Session, Run Storage format, Assembly Plan, Sampling, workflows, backend, and GPU were unchanged. Audio multi-reference, Issue #13, A5, and A8b remain separate/HOLD. No commit, push, tag, release, GitHub, or Civitai action was performed.
|
||
|
||
## V3.8 Non-destructive Production Phase N1 PASS (2026-09-03)
|
||
|
||
- The normal `H3ContinuumSamplerV38` Production view now adds a frontend-only Run Plan plus `Continue`, `Regenerate Current`, and `Finish Remaining` shortcuts. Each shortcut selects the existing Review Each Chunk / Review Action contract, enables Save + Auto Resume, and resets manual Regenerate From to Auto; it does not queue automatically or introduce a second execution path.
|
||
- The Run Plan reports only current widget intent: next Queue action, Run Storage mode, requested Regenerate From, and requested Variation Nonce. Current physical group and Accepted/Reused/Regenerated/Pending results are not inferred in frontend; the existing post-Queue Sampler Status remains authoritative for those Run Storage facts.
|
||
- All four helper widgets are transient and are removed around workflow serialize/configure. Basic hides them; Production reveals them. Existing 24 serialized widget values/order, inputs, sockets, queue keys, and saved-workflow reload behavior remain unchanged. One-shot Regenerate/Finish still resets to Continue only after an accepted Queue through the existing `afterQueued` lifecycle.
|
||
- Continue maps to implicit acceptance plus one next physical group, Regenerate Current maps to Smart Regenerate of the latest review unit with backend nonce advance, and Finish Remaining reuses the accepted prefix then removes the one-group cap. The existing resolver/Run Storage tests retain Terminal Merge atomicity and restart/reuse behavior.
|
||
- Validation passed focused `105`, full pytest `1048`, JavaScript syntax `7/7`, development Manifest `247/247`, Registry Manifest `138/138`, runtime verifier/public export `6/6`, source/runtime frontend SHA `1/1`, ten protected backend/schema/contract files unchanged from the pre-N1 snapshot, and `git diff --check`.
|
||
- Source snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-n1-main-production-ux-filtered-20260903_012203-0a7c670` (`314/314`, mismatch zero). Runtime snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v38-n1-runtime-w-20260903_013138` (`1/1`, exact). Only `web/project_id.js` was synchronized to ComfyUI_W; no backend, port 8190, port 8188, live Frontend, or GPU Sampling was operated.
|
||
- Formal result: **N1 PASS**. Production Python Schema, Run Storage/State/Session formats, Assembly Plan, Sampling, Seed, SIGMAS, Terminal Merge, workflows, package surface, and public Node count are unchanged. Render History / Branch / Take and a live structured timeline remain Phase N2. Audio multi-reference, Issue #13, A5, and A8b remain separate/HOLD as previously fixed. No commit, push, tag, release, GitHub, or Civitai action was performed.
|
||
|
||
## V3.7.0 Civitai Reproduction / V3.7.1 Hotfix Gate (2026-09-03)
|
||
|
||
- Ran the released `v3.7.0` tag commit `fe4ff9c20c2cc8bb375625d1534f5673a737d1be` in an isolated ComfyUI 0.34.2 base directory on dedicated `127.0.0.6:8190`; only the historical Continuum package and an evidence probe were loaded. External Spectrum/Sol/FBC/Turbo/LoRA/model wrappers were absent.
|
||
- The version-exact 2x5 Dialogue A/B used the same prompt, seed `3802902`, model, VAEs, simple scheduler, eight steps, 384x384, List mode, Balanced 22, Audio Continuity ON, Standard, Run Storage OFF, and Seam OFF as the accepted V3.8 Gate. V3.7.0 and V3.8 retained all eight words (`Alpha` through `Hotel`) and matched bit-exact for Video prefix, Audio prefix, raw Group 1/2 Audio, and Finalize Audio. CASE A and CASE B were not reproduced; V3.8 did not uniquely repair this controlled case.
|
||
- Wrapper-free V3.7.0 Reference Audio completed without matrix mismatch or packed-row error. Fixed 3x5 preserved both latent prefixes and showed no character identity reset. Diagnostic Audio Seam OFF exposed decoded boundary jumps `0.094740 / 0.415353`; the same cached raw latents with the public default Audio Seam Auto reduced them to `0.042853 / 0.034786`. A semantic music-phrase restart was not proven and remains inconclusive.
|
||
- Per the approved rule, no V3.7.1 candidate or speculative Production fix was created. Historical pytest passed `632`; V3.7 runtime verification and native PackedLayout passed; tag extraction matched `226/226` Git blobs. Current source matched the pre-Gate snapshot `314/314` before this state entry. Peak VRAM was `15.519 GiB`.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v370-civitai-reproduction-hotfix-20260903_003300`. Snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-v370-civitai-reproduction-hotfix-gate-filtered-20260903_003041-0a7c670`. Dedicated 8190 is stopped; 8188 and Frontend were untouched. No commit, push, tag, release, or external post was performed. Reopen only with reporter-exact workflow/API prompt, seed, traceback, model/LoRA, and wrapper/plugin versions.
|
||
|
||
## V3.8 Emergency Dialogue / Audio Release Blocker Gate PASS (2026-09-03)
|
||
|
||
- The dedicated V3.8 `2x5`, 24 fps, 384x384, Balanced 22, Audio Continuity ON, Standard, List-prompt GPU Gate did not reproduce the Civitai dialogue truncation report. Raw Group 2's first 22f/0.9167s were the protected prior-audio prefix; Video and Audio latent prefixes were bit-exact. The raw retained suffix and Finalize Chunk 2 both contained all eight requested words (`Alpha` through `Hotel`), and the Finalize segment matched raw Group 2 from sample 29,333 bit-exact. CASE A physical-audio loss and CASE B semantic pre-roll loss were both not reproduced.
|
||
- The same wrapper-free V3.8 graph passed with Reference Audio connected through Core `LoadAudio`; no matrix mismatch, packed-row failure, NaN/Inf, OOM, Decode, or Assembly error occurred. The basic Continuum/Core Reference Audio path is healthy. The reported hard error remains most consistent with an external/obsolete wrapper such as the known pre-fix Turbo packed-row implementation; do not change Continuum tensor shapes without reporter-exact traceback/version evidence.
|
||
- The same subject and studio remained visually continuous across the tested boundary with no major seam. Music reset was not evaluated because the minimum dialogue prompt contained no music; no additional semantic GPU Matrix was opened.
|
||
- No Production fix or candidate resampling was justified. Production code, public schema/widgets/sockets, Sampling, Run Storage, State/Session, Assembly Plan, workflows, and runtime package are unchanged. Full pytest passed `1047`; focused static tests `147`; Manifest `247/247`; runtime verifier, JavaScript `7/7`, critical source/runtime SHA `10/10`, and `git diff --check` passed. Peak VRAM was `15.807 GiB`.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-emergency-dialogue-audio-release-blocker-20260902_234859`. Required pre-fix snapshot: `pre-v38-emergency-dialogue-audio-release-blocker-fix-filtered-20260902_233838-0a7c670` (`314/314`, mismatch zero). Dedicated 8190 and its temporary probe are stopped/removed; 8188 and Frontend were untouched.
|
||
- Formal state: **Emergency Dialogue/Audio Release Blocker Gate = PASS / Production fix not required / prior Package Staging Re-Gate remains PASS / Release & Registry readiness = GO on current evidence.** Reopen only with the exact reporter workflow, traceback, and dependency versions. No commit, push, tag, release, or external post was performed.
|
||
|
||
## V3.8 Public Surface Cleanup complete (2026-09-02)
|
||
|
||
- The supported/searchable V3.8 surface is now exactly six IDs: `H3ContinuumSamplerV38`, `H3ContinuumAssembleSeamV35` (displayed as `H3 Continuum Finalize`), `H3EasyLoadImage`, `H3EasyLoadAudio`, `H3ContinuumLoadVideo`, and `H3ContinuumSecondPassV35`. Legacy, Labs, and diagnostic implementations remain in source where required but are not merged into the package root mappings.
|
||
- Main presentation uses a frontend-only `H3 Continuum View` property with `Basic` and `Production`. Switching views changes visibility only: Python input keys/order, widget count/values, sockets, prompt payload fields, Sampling behavior, Run Storage, State/Session, and Assembly contracts are unchanged.
|
||
- The Registry package exposes only `examples/workflows/MiniMax_H3_Continuum_V38.json`; historical workflows, tests, tools, diagnostics, caches, and bytecode are excluded by `.comfyignore`. V3.7/V3.6 remain available through their existing tags; older V3.4/V3.5 archival packaging is deferred to release preparation.
|
||
- Workflows containing non-exported IDs intentionally load as unknown nodes in V3.8 and require the corresponding historical release. The source classes/modules were preserved to avoid breaking imports used by the six supported nodes.
|
||
- Validation passed: focused public-surface tests `18`, full pytest `1046`, source runtime verifier, installed-runtime export/display `6/6`, JavaScript syntax, `git diff --check`, Manifest `245/245`, and source/runtime SHA `8/8`. The standard `tools/validate.ps1` reached pytest but its default cache/global temp paths hit the known Windows ACL denial; the same full suite passed with cache disabled and a dedicated writable `--basetemp`.
|
||
- Source snapshot: `pre-v38-public-surface-cleanup-excluding-cache-20260902_210308-0a7c670` (`309/309`, SHA mismatch zero). Exact pre-sync runtime snapshot: `pre-v38-public-surface-runtime-w-20260902_222000` (`7/7`, SHA mismatch zero; the V3.8 workflow was absent). Only eight public-surface runtime files were synchronized to ComfyUI_W; no backend, 8188, or Frontend operation was performed.
|
||
- Commit, push, tag, release, and GPU Sampling were not performed. A5 and A8b remain HOLD; no promotion or behavior claim is implied by this cleanup.
|
||
- `H3情報チェック` formally accepted **V3.8 Public Surface Cleanup = PASS** with no mandatory code correction and no additional GPU Gate. Release/Registry readiness remains **HOLD** until a later package-staging dry-run verifies the actual `.comfyignore` archive contents and imports that staged package from an isolated directory with the same six-node surface.
|
||
|
||
## V3.8 Product Direction Freeze and public-surface read-only audit (2026-09-02)
|
||
|
||
- The formal product definition is: **H3 Continuum is a Production Sampler that generates, reviews, partially regenerates, and resumes long-form video without restarting the entire work.** The product has two layers only: `Main / Production` and `Advanced`. Its three pillars are `Continuation Quality`, `Non-destructive Production`, and `Open Integration`.
|
||
- `Main / Production` is simple by default and expands to Production controls; there is no separate future Easy product path. `Advanced` owns Second Pass, Selective/Temporal Refine, external integration, and diagnostics. Existing unreleased Easy classes may remain temporarily in source but carry no V3.8 public compatibility obligation.
|
||
- V3.8 is the current supported product. V3.7 and earlier are historical downloads through GitHub Release/tag, not public Node-ID or saved-workflow compatibility requirements inside V3.8. Local tags `v3.7.0` and `v3.6.0` are present; older archival packaging remains a release-preparation task.
|
||
- The audited target launch surface is six searchable nodes: `H3ContinuumSamplerV38`, the existing `H3ContinuumAssembleSeamV35` implementation presented later as Finalize, `H3EasyLoadImage`, `H3EasyLoadAudio`, `H3ContinuumLoadVideo`, and `H3ContinuumSecondPassV35`. Display-name and registration changes were not made in this audit.
|
||
- Current source registers `40` IDs. The target classification is `6 Keep Public`, `3 Labs`, `2 Remove Registration`, and `29 Legacy Release Only`. The separately installed `H3_Continuum_GPU_Diagnostics` runtime pack adds three development nodes and is excluded from the V3.8 release surface.
|
||
- Main UI/Schema verdict is **UI-only GO with one boundary**. Existing widgets can be hidden or expanded without changing Python keys/order because the frontend already keeps hidden widgets in place. A true serialized single `Duration` input does not exist in Main V3.8; V3.8 currently uses `chunks` plus `chunk_seconds`. The initial UI implementation should retain those backend fields and may show a computed total rather than migrate schema.
|
||
- Non-destructive backend support already exists for Review Each Chunk, implicit accept through `Continue / Next`, Regenerate Current, Smart Regenerate, Regenerate From, Finish Remaining, automatic resume, prefix reuse, branch boundary, and effective nonce. The main missing product surface is structured Render History/Take selection and a clear timeline display of accepted/reused/pending state.
|
||
- Hi-Res Fix is outside the new standard workflow. Assembly remains mandatory while Seam correction is optional; the current assembler can be presented as Finalize without moving Core Video/Audio Decode into the Sampler. The existing Node ID should be reused for the first UI implementation rather than adding another assembler class.
|
||
- P1 readiness evidence is retained while P1a FBC and P2 Sol-Attn are `Deferred`. A5 and A8b remain `HOLD`; Issue #13 reporter-exact R3 remains the highest Continuation Quality investigation when the reporter workflow becomes available. Production code, runtime, backend, Frontend, schema, Sampling, Run Storage, State/Session, and Assembly were not changed by this docs/audit phase.
|
||
- Audit evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-product-direction-freeze-20260902_202941`. Pre-change snapshot: `pre-v38-product-direction-public-surface-audit-excluding-cache-20260902_202606-0a7c670` (`309/309`, SHA-256 mismatch zero). The standard helper first encountered the known `.pytest_cache` ACL denial; the accepted fallback excludes ignored cache/bytecode only.
|
||
- `H3情報チェック` accepted the audit and marked `Public Surface Cleanup` GO. Mandatory boundaries are: allowlist only the exported class/display mappings; retain legacy imports/classes used by V3.8; document that workflows containing non-exported IDs become unknown nodes and require their historical release; keep Basic/Production presentation-only without changing hidden values, backend kwargs, Sampling identity, or Run Storage identity; keep production helpers, metadata, notices, current workflow, and runtime-imported legacy modules in the package. No implementation was started.
|
||
|
||
## Current HOLD release conditions after A8a (2026-09-02)
|
||
|
||
- Current accepted baseline is **A8a Adaptive Continuity Observer = GPU Experimental PASS**: full pytest `1041`, Manifest `242/242`, source/runtime SHA `2/2`, five GPU Gates passed, and every record fixes `execution_applied=false`. Production, Sampling, A7b, Run Storage, State, Session, and Assembly remain unchanged.
|
||
- **A8b execution policy remains HOLD.** Release requires one explicitly defined intervention type only; no combined context/transport/mask/latent/conditioning change; Seed/SIGMAS/physical grouping unchanged unless separately approved; and no reuse of the failed Issue #13 13f, 22f/75%-preserved, Soft Re-anchor, or Production Fixed Prefix approaches.
|
||
- Before A8b GPU Sampling, either reporter-exact Issue #13 R3 must complete or static plus replay evidence must prove that the intervention is independent of the recursive-feedback path. The A8a recommendation must then pass Shadow/Replay validation with no unnecessary intervention, Terminal Merge misclassification, or Run Storage replay difference.
|
||
- Any initial A8b path must be Experimental and Default OFF. Existing Production stays bit-exact; public schema change is a separate approval; external Sage/Sol/Spectrum wrappers remain intact. These are the minimum conditions for **A8b Experimental GPU Gate GO**, not Production promotion.
|
||
- **A5 Reference Video Influence Schedule/Fade remains HOLD.** Reporter-exact Issue #13 R3 must classify Reference Video as unrelated, amplification-only with a defined safe range, or necessary/strong cause candidate. Unrelated permits A5 GO; amplification-only permits a bounded Experimental GO; necessary/strong cause retains HOLD.
|
||
- A5 OFF must be bit-exact with Production. The initial A5 implementation may change Reference Video influence only and must not change Video continuation, Audio, Seed, SIGMAS, Terminal Merge, Assembly, or Run Storage. A5 and A8b must not be implemented concurrently.
|
||
- A5 requires at least `3x5` and a long-form continuation GPU Gate covering trajectory, identity, saturation, contrast, sharpness, and highlight drift. Short-form evidence alone cannot promote it.
|
||
- Formal next-state graph: `A8a PASS -> A8b HOLD (single policy + Issue #13 independence/R3 + Shadow/Replay)`, and `Issue #13 reporter workflow -> R3 -> Reference Video causal classification -> A5 HOLD release decision`.
|
||
- Pre-documentation snapshot: `pre-v38-a8b-a5-hold-release-contract-20260902_191322-0a7c670` (`309` files, mismatch zero).
|
||
|
||
## Advanced Phase A7b Memory Action Policy v1 GPU Experimental PASS (2026-09-02)
|
||
|
||
- A7b v1 is deliberately limited to an explicit opt-in Reference downgrade policy. The existing public `H3 Continuum Sampler V3.8` schema and default behavior are unchanged; the new Experimental policy node and Experimental sampler variant are the only activation path. `Disabled` delegates to the existing V3.8 sampler without changing its arguments or outputs.
|
||
- Policy contract v1 supports only `Reduce References One Step`: Reference Images map `Max Identity -> Match Output`, and Video Guide maps `Match Output -> Balanced 0.6 MP -> Efficient 0.4 MP`. Already-lower settings and disconnected inputs stay unchanged. Malformed or unsupported contracts fail soft to `Disabled`.
|
||
- Automatic MODEL unload and attention/backend switching are advisory-only and are never executed. Same-run actual packed-row observations do not trigger policy changes. External MODEL wrappers, including the active Sage route, remain untouched. Production Sampling, PackedLayout, extra conditions, seed, SIGMAS, physical-group order, Terminal atomicity, Assembly semantics, and package version are unchanged.
|
||
- CPU/static validation passed focused V3.8/schema/workflow `49`, A7a/Reference/V3.8 related `120`, full pytest `1027`, Manifest `240/240`, runtime verifier, Python syntax, JavaScript `7/7`, and `git diff --check`. `H3情報チェック` accepted the CPU/static Gate and approved the five-case GPU Matrix without a mandatory correction.
|
||
- Dedicated ComfyUI_W/Core `0.34.2` GPU testing passed G1-G5. G1 `Disabled` was bit-exact against the accepted A7a Production baseline for Video/Audio latent, decoded RGB/PCM, normalized Assembly, sampler/SIGMAS links, and seed inputs. G2 Reference Image 1 reduced actual rows `15,144 -> 13,128`; G3 five-frame Video Guide reduced `14,230 -> 13,639`; G4 three Reference Images reduced `20,546 -> 14,498`; G5 124-frame Video Guide reduced `37,790 -> 29,319`.
|
||
- Every active policy case preserved Reference count/frame count, fixed seed, SIGMAS, physical/Assembly semantics, and the external MODEL wrapper while changing effective Reference geometry, conditioning hash, output variation, and Run Storage global identity as required. No unload or backend mutation occurred. No OOM, NaN/Inf, Decode, Assembly, duration, or PCM failure occurred.
|
||
- GPU API seconds were G1 `102.649`, G2 `117.286 / 100.437`, G3 `112.951 / 108.319`, G4 `160.027 / 99.069`, and G5 `359.271 / 257.536` (`1,417.546 s` total). Peak VRAM was `15.516 GiB`, peak backend Working Set `45.724 GiB`, and minimum available RAM `5.479 GiB`; 16GB passed with small headroom.
|
||
- The observed packed-row reductions and shorter warm second-run API times are evidence for these fixed cases only, not a guaranteed VRAM or speed reduction contract. Cache warmth, backend state, model, geometry, and conditioning can change the performance result.
|
||
- Post-GPU validation again passed full pytest `1027`, Manifest `240/240`, runtime verifier, Python/evidence-runner syntax, JavaScript `7/7`, source/runtime SHA `3/3`, and `git diff --check`. Source snapshot: `pre-v38-a7b-action-policy-excluding-cache-20260902_143407-0a7c670` (`303` files, mismatch zero). Exact pre-sync runtime snapshot: `pre-v38-a7b-gpu-runtime-w-20260902_154625` (`2/2` existing files, mismatch zero; the new policy module was absent). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a7b-memory-action-gpu-matrix-20260902_154825\summary.json`.
|
||
- Queue was empty before stopping only the command-line/parent-verified dedicated backend PID `36400` and launcher PID `38804`; port 8190 is closed. The two generated A7b Reference test videos were removed from ComfyUI_W input after the backend stopped; the original source image remains. Port 8188 and the Frontend were not operated. Commit, push, tag, and release were not performed. A5 remains HOLD.
|
||
- `H3情報チェック` formally accepted A7b v1 as GPU Experimental PASS with no mandatory correction, additional GPU Gate, or re-Sampling. A standalone `Match Output -> Balanced 0.6 MP` GPU case is optional rather than required because the resolver chain and CPU tests already cover it.
|
||
- Accepted source snapshot: `v38-a7b-gpu-experimental-pass-accepted-excluding-cache-20260902_162542-0a7c670` (`306` files, mismatch zero; excludes `.git`, `.pytest_cache`, `__pycache__`, and `.pyc`; the inherited `.pytest_cache` ACL emitted the known enumeration warning only).
|
||
- Formal state: **A7b v1 Memory Action Policy = GPU Experimental PASS / explicit opt-in Reference downgrade only / default Disabled bit-exact / unload advisory-only / backend switch advisory-only / external wrappers and Production unchanged / A5 HOLD**.
|
||
|
||
## Advanced Phase A6 Temporal Window Refine GPU Experimental PASS (2026-09-02)
|
||
|
||
- A6 is independent of the held A5 Reference Video Influence Schedule. The existing public Experimental Selective Second Pass node now adds `Refine Scope = Time Window` with `Window Start (sec)` / `Window End (sec)` appended after the A4 widgets; `All` remains the default and saved A3/A4 workflows that lack the new tail values restore `0.0 / 5.0` without changing prior behavior.
|
||
- New pure `v3/refine_window.py` contract v1 maps a half-open interval on the **final visible output timeline** through exact-duration trim/pad/final-frame preservation, natural logical ranges, physical-group-local frame ranges, H3 `17k+5` Video slots, and Audio ticks. Continuation trim prefixes are always excluded. Selecting either logical half of a Terminal Merge keeps its physical group atomic.
|
||
- Non-intersecting physical groups are skipped before prompt encode, refine-context adaptation, MODEL clone, seed derivation, and Sampling, preserving their original Video/Audio LATENT objects. Intersecting groups use temporal denoise masks and then explicitly restore every out-of-window Video slot and Audio tick from the input tensor. This is group-level compute pruning only; A6 makes no intra-group Sampling-cost reduction claim.
|
||
- `Video Only`, `Audio Only`, and `Video + Audio` window contracts passed mocked CPU/static checks for masks, stream selection, prefix protection, terminal atomicity, final-frame timeline inversion, exact outside-window restoration, group skipping, unchanged Assembly Plan fields, and legacy no-window dispatch. V3.5, Production Sampling, V3.8 Sampler, RefineSchedule v1, Run Storage, Frontend, public workflows, and package version are unchanged.
|
||
- The dedicated 576x576 fixed-seed GPU Matrix passed 5/5 on ComfyUI_W/Core `0.34.2`: G1 no-window/All Video Only was bit-exact with legacy V3.5; G2 Video Only `[1,3)` changed only selected Video slots and preserved Audio exactly; G3 Video+Audio `[1,3)` changed only selected AV ranges; G4 `3x5` `[4,6)` sampled Groups 1/2 only and exactly reused Group 3; G5 FL2VA Terminal `[2,3]` `[5.5,8)` reused Group 1 and sampled the complete terminal pair once. Every selected group kept the same absolute seed and SIGMAS as its target-matched All baseline. Window state remained absent from Assembly Plan and Run Storage.
|
||
- One initial G2 evidence STOP was a runner-only target mismatch: Video Only seed/field set was compared to a Video+Audio baseline. G2's actual tensor, mask, prefix, media, and schedule checks had all passed. The runner was corrected to use target-specific namespaces/contracts; successful G1/G2 evidence was revalidated without resampling, and G3-G5 then passed. Production code was not changed.
|
||
- GPU API seconds were `210.845 / 90.768 / 92.013 / 378.361 / 478.395`. Peak device use was `15.514 GiB`, peak backend Working Set `42.004 GiB`, and minimum available RAM `3.403 GiB`; 16GB passes with small VRAM headroom. All outputs were 24 fps, exact 5/15 seconds, 32 kHz stereo, finite and boundary-clean, with no OOM, NaN/Inf, duration, Decode, or Assembly error.
|
||
- Post-GPU validation passed focused `15`, related A3/A4/V3.5 `82`, full pytest `1004`, Manifest `236/236`, runtime verifier, Python syntax including the evidence runner, JavaScript `7/7`, source/runtime SHA `5/5`, and `git diff --check`. The complete suite used the ComfyUI_W Python with dedicated writable `--basetemp` because the ordinary `%TEMP%\\pytest-of-links` has an inherited ACL denial.
|
||
- Snapshots: source `pre-v38-a6-temporal-window-contract-excluding-cache-20260902_115158-0a7c670` (`297` files, mismatch zero) and exact pre-sync runtime `pre-v38-a6-gpu-runtime-w-20260902_122034` (`4` existing files, mismatch zero; `v3/refine_window.py` was absent). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a6-temporal-window-refine-gpu-matrix-20260902_122356\summary.json`.
|
||
- Accepted source snapshot: `v38-a6-gpu-experimental-pass-accepted-excluding-cache-20260902_125452-0a7c670` (`300` files, mismatch zero; excludes `.git`, `.pytest_cache`, `__pycache__`, and `.pyc`).
|
||
- Queue was empty before stopping only dedicated 8190 backend PID `12808` and launcher PID `13648`; port 8190 has no listener. Port 8188 and Frontend were not operated. Commit, push, tag, and release were not performed. A5 remains HOLD; Audio-targeted modes and the public Selective node remain Experimental/Production HOLD; A7 has not started.
|
||
- `H3情報チェック` formally accepted A6 with no mandatory correction or additional GPU Gate. For G4, `[4,6)` intersected physical Groups 1/2 only; Group 3 stayed fully outside the window and was reused exactly.
|
||
- Formal state: **A6 Temporal Window Refine = GPU Experimental PASS / final-visible-timeline scoped / group-level skip supported / intra-group compute pruning not claimed / outside-window output explicitly restored bit-exact / Terminal Merge atomic / Production unchanged / A5 HOLD**.
|
||
|
||
## Issue #13 R2.8 Continuation Depth / Strength Gate FAIL (2026-09-02)
|
||
|
||
- R2.8 audited the actual masked AV continuation contract before GPU work. `Balanced 22` selects the last 22 pixel frames on H3's `17k+5` grid, represented as 7 Video temporal latent slots and 37 Audio ticks. The accepted tails are copied into the next target latent under nested denoise masks; accepted Video/Audio prefixes are restored bit-exact after Sampling. Normal recursive continuation does not add a decoded-frame VAE round trip, and there is no separate public H3 continuation-strength scalar.
|
||
- Baseline A reused accepted R2.6-A. B1 used the approved H3-valid effective 13f mapping: last 4 of 7 Video slots active, older 3 prefix slots fully denoised, Audio 37T unchanged. It failed by multiple metrics and visual review: Saturation improved `13.66%` and skin-highlight `13.82%`, while Contrast worsened `331.02%`, Sharpness `56.98%`, Gradient `7.52%`, and ordinary highlight clipping `132.45%`; boundary-flow ratio was `1.025`. Per the stop rule, effective 9f was not run.
|
||
- C1 used only Core's supported fractional denoise mask on the 22f Video protected prefix: mask `0.25` = 25% denoise / 75% preserved. Audio remained fully protected. Sharpness improved `27.47%` and skin-highlight `61.55%`, but Saturation worsened `39.10%`, Contrast `229.74%`, Gradient `15.98%`, and ordinary highlights `50.68%`; the Group 1→6 face-structure proxy fell to `0.484`, matching visible later face/pose/composition drift. C1 failed. No lower strength was tested.
|
||
- B and C were not independently promising, so the combined condition was not run. The aggregate screening means (`-83.42% / -41.08%`) remain diagnostic only; the formal FAIL is based on consistent individual-metric and visual evidence, not a composite product score.
|
||
- Both GPU candidates preserved Video/Audio prefix integrity `5/5`, six physical groups/six Sampling calls, prompt/SIGMAS contracts, 720 frames, exact 30 seconds, 24 fps, 576x864, and 32 kHz stereo. No OOM, NaN/Inf, decode, assembly, duration, or Queue error occurred. API totals were `897.447 / 864.321 s`; peak device memory was `14.854 GiB` and minimum available RAM `0.639 GiB`.
|
||
- Validation passed focused `21`, full pytest `989`, Manifest `234/234`, runtime verifier, Python syntax, JavaScript `7/7`, source/runtime SHA `5/5`, and `git diff --check`. Dedicated 8190 was stopped after Queue empty; 8188 and Frontend were not operated. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r28-20260902_022000\ISSUE13_R28_CONTINUATION_CONDITIONING_REPORT.md`.
|
||
- The R2.8 implementation is diagnostic-only: the minimal private hook defaults to `None`, is unreachable from Production public schemas/workflows, and leaves default behavior unchanged. Production defaults, State/Session, V3.8 schema, Assembly, Run Storage, Frontend, version, commit, push, tag, release, and GitHub Issue were unchanged.
|
||
- `H3情報チェック` formally accepted R2.8 FAIL with no mandatory correction or additional GPU Gate. Formal state: **R2.8 complete / effective Video history 13f FAIL / 9f not run / 22f Video protected-prefix 25% denoise mask (75% preserved) FAIL / combination not run / recursive continuation feedback remains supported as the dominant amplification pathway / Production behavior unchanged / A5 HOLD**.
|
||
- Accepted completion snapshot: `issue13-r28-depth-strength-fail-accepted-excluding-cache-20260902_025646-0a7c670` (`297` files, mismatch zero; excludes `.git`, `.pytest_cache`, `__pycache__`, and `.pyc`).
|
||
|
||
## Issue #13 R2.7 one-shot Soft Re-anchor mitigation FAIL (2026-09-02)
|
||
|
||
- R2.7 compared normal recursive continuation against one diagnostic-only 50% low-frequency Y/U/V appearance re-anchor on the temporary Group 4 Video conditioning context. Both used the accepted R2.6 close-up 576x864 `6x5` Turbo same-prompt contract with seed `131325`, Euler/simple 8 evaluations, Balanced 22 masked AV continuation, and Audio Continuity ON.
|
||
- The re-anchor contract itself passed: exactly one event, Group 1 appearance reference, current Group 3 tail retained, target Group 4, Video context/geometry shape unchanged, source/State inputs not mutated, and Audio object passed through. Groups 1–3 Video/Audio outputs and all six seed/SIGMAS/noise/prompt signatures matched A exactly.
|
||
- Mitigation was mixed rather than compound: Saturation buildup fell `32.1%` and ordinary highlight clipping reversed, but luma std worsened `252.7%`, Sharpness worsened `11.4%`, Gradient worsened `1.1%`, and the skin-highlight proxy worsened `5.7%`. Alternative `37.2% / 29.9% / 23.9%` aggregate calculations are diagnostic appendix only because highlight reversal dominates them; **no valid composite mitigation score was established**.
|
||
- No major visible seam, identity collapse, pose reset, or trajectory break occurred. However, Group 3→4 flow discontinuity increased `0.091975 -> 0.107117` (`+16.5%`), so continuity impact is not cleanly neutral. Both outputs remained exact 720 frames / 30 seconds / 24 fps with 32 kHz stereo Audio; no OOM, NaN/Inf, decode, assembly, or duration error occurred.
|
||
- API seconds were `939.830 / 884.677`; B ran warm second, so the faster total is not an overhead estimate. Peak VRAM was `14.934 GiB`, peak backend Working Set `42.166 GiB`, and minimum available RAM `2.867 GiB`; 16GB passed with small headroom.
|
||
- Validation passed focused `10`, full pytest `983`, Manifest `234/234`, runtime verifier, Python `5/5`, JavaScript `7/7`, A4 Production SHA `4/4`, diagnostic SHA `2/2`, and `git diff --check`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r27-20260902_005002\ISSUE13_R27_SOFT_REANCHOR_AB_REPORT.md`.
|
||
- Accepted source snapshot: `issue13-r27-soft-reanchor-fail-accepted-excluding-cache-20260902_013547-0a7c670` (`293` files, mismatch zero; excludes `.git`, `.pytest_cache`, `__pycache__`, and `.pyc`).
|
||
- The pinned `H3情報チェック` review formally accepted the FAIL classification with no mandatory correction or additional GPU Gate. It fixed the Video wording to `recursive-prefix parity: intentionally not applicable after the re-anchor event`; this is a designed diagnostic intervention, not prefix corruption.
|
||
- Formal state: **Issue #13 R2.7 complete / 50% single-event appearance Soft Re-anchor FAIL as a mitigation candidate / partial saturation and highlight signal only / sharpness, gradient, skin-highlight and flow composite not improved / continuity side effects not neutral / diagnostic intervention behaved as designed / Production unchanged / adaptive or repeated re-anchor not promoted / A5 HOLD**. R2.6 remains unchanged: recursive continuation feedback is strongly supported as the dominant amplification pathway, while global decoded appearance statistics alone are insufficient to suppress it. No alternate strength, Adaptive/multiple re-anchor, Production/UI work, commit, push, tag, release, or Issue mutation was performed.
|
||
|
||
## Issue #13 R2.6 dominant amplification pathway strongly supported (2026-09-02)
|
||
|
||
- R2.6 completed the R2.5 fallback confirmation with two matched close-up `6x5` Turbo runs at 576x864, fixed seed `131325`, Euler/simple 8 evaluations, Balanced 22-frame masked AV continuation, Audio Continuity ON, and the exact same prompt in all six chunks. A used normal recursive continuation; B used diagnostic teacher forcing that captured the Group 1 Video/Audio tail and reused it for Groups 2–6.
|
||
- Same-prompt recursive A still showed decoded Chunk 1-to-6 drift: saturation `+0.020101`, sharpness `+171.377`, gradient `+0.041037`, and highlight clipping `+0.004487`. Prompt schedule is therefore not a necessary cause.
|
||
- Fixed Prefix B reduced those positive deltas to `+0.004639 / +56.616 / +0.011058 / +0.001967`, reductions of `76.9% / 67.0% / 73.1% / 56.2%`. **Recursive continuation feedback is strongly supported as the dominant amplification pathway under this controlled condition.** Suppression was not complete, so the smaller initiating Sampling/conditioning bias remains unresolved.
|
||
- All six A/B groups matched derived seed, SIGMAS, Video/Audio noise, prompt embedding, sampler, and model options. Groups 1–2 matched input/output exactly; Groups 3–6 diverged only after the selected continuation source diverged. All five B boundaries used the fixed Video/Audio source correctly. B's normal recursive prefix-pair result is intentionally false from Group 3 because teacher forcing does not use the immediate previous tail; this is not Assembly corruption.
|
||
- Both GPU runs produced exact 30-second, 720-frame, 24-fps outputs without OOM, NaN/Inf, decode, or assembly errors. API seconds were `915.064 / 881.397`. Dedicated 8190 is stopped; 8188 and the Frontend were not operated.
|
||
- Diagnostic-only changes remain limited to `tools/gpu_diagnostic_node/nodes.py`, `tools/issue13_r25_gate_runner.py`, `tools/issue13_r2_analyze.py`, and `tests/test_issue13_r25_tools.py`. Production Sampling, Conditioning, Run Storage, Assembly, public schema, Frontend, workflows, and package version are unchanged.
|
||
- Validation passed focused `6`, full pytest `979`, Manifest `234/234`, runtime verifier, Python `4/4`, JavaScript `7/7`, A4 Production source/runtime SHA `4/4`, diagnostic source/runtime SHA `1/1`, and `git diff --check`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r26-20260901_234141\ISSUE13_R26_FIXED_PROMPT_TEACHER_FORCING_REPORT.md`.
|
||
- Accepted completed source snapshot: `issue13-r26-dominant-feedback-complete-excluding-pytest-cache-20260902_002924-0a7c670` (`833` files, mismatch zero; only `.git` and the inherited inaccessible `.pytest_cache` excluded).
|
||
- The pinned `H3情報チェック` review formally accepted R2.6 as PASS with no mandatory correction or additional GPU Gate. It required the technically narrower phrase `strongly supported as the dominant amplification pathway`, retained A5 HOLD, and classified any initiating-bias trace as a separate Research Gate rather than a Production-fix prerequisite.
|
||
- Recorded limits: Fixed Prefix is diagnostic teacher forcing, not a Production mode; contrast did not rise in the same-prompt pair; recursive feedback is not claimed as the only cause; Motion LoRA, model/Core defect, and conditioning bias contributions are not fully excluded; and the 576x864, one-seed, `6x5`, Turbo result is not generalized to every setting.
|
||
- Formal wording: **Issue #13 R2.6 complete / same-prompt degradation persists / fixed-prefix substantially suppresses degradation / recursive continuation feedback strongly supported as the dominant amplification pathway / initiating root cause unresolved / Production unchanged / A5 HOLD**. Commit, push, tag, release, and GitHub Issue mutation were not performed.
|
||
|
||
## Issue #13 R2.5 Cause Isolation PASS / leading classification C (2026-09-01)
|
||
|
||
- The staged close-up GPU Gate completed at 576x864, fixed seed `131325`, Euler/simple, Balanced 22-frame masked AV continuation, and Audio Continuity ON. Core screening used the official MiniMax H3 FL2V Turbo 4-step v1.0 LoRA with 8 scheduler evaluations and slightly evolving per-chunk prompts under fixed identity/camera/exposure constraints.
|
||
- `3x5` and `3x10` did not reproduce the full oil-film/plasticization composite. `6x5` showed the later-group combination of contrast, highlight clipping, edge hardness, and facial gloss. The leading classification is **C: continuation-count / cumulative-feedback dependent**, not chunk-duration dependent. It is not yet a final root-cause proof because the approved prompt schedule also grows from three to six distinct chunk prompts.
|
||
- In `6x5`, decoded center contrast progressed `0.8978, 0.8963, 0.9073, 0.9060, 0.9267, 0.9165`, highlight clipping `0.0035, 0.0059, 0.0074, 0.0071, 0.0168, 0.0135`, and sharpness `401.6 -> 531.3`. The decoded continuation-input proxy contrast progressed `0.8940 -> 0.8959 -> 0.9029 -> 0.9131 -> 0.9219`, so the later input already carries the change. Final accepted Video/Audio prefixes remained bit-exact for all `5/5` boundaries; prefix/Assembly corruption was not observed.
|
||
- The staged stop rule omitted `4x10/5x10/6x10`. Motion LoRA isolation then compared no LoRA versus H3 Motion Booster V0.2 + BUNNY Motion Continuity Repair v1.0 at identical normal 20-step Euler/simple conditions. Both LoRAs were actually applied (`208 patches`) and changed motion, but neither the OFF/OFF nor ON/ON 3x5 run reproduced the composite degradation. The pair is not a necessary cause because `6x5` reproduced without either Motion LoRA; single-LoRA cases were not required.
|
||
- GPU/API runs all completed without OOM, NaN/Inf, decode, or assembly error. API seconds were `482.119 / 1038.856 / 939.591 / 903.117 / 975.867` for `3x5 Turbo / 3x10 Turbo / 6x5 Turbo / Motion OFF / Motion BOTH`. Captured CUDA allocator peak-reserved evidence was `3.281-6.531 GiB`; this is not a device-wide peak-VRAM reading.
|
||
- Diagnostic-only changes are limited to `tools/gpu_diagnostic_node/nodes.py`, `tools/issue13_r25_gate_runner.py`, `tools/issue13_r2_analyze.py`, and `tests/test_issue13_r25_tools.py`. Production Sampling, Conditioning, Run Storage, Assembly Plan schema, Frontend, workflows, and package version are unchanged. Focused tests passed `9`; full source regression passed `977`; runtime verifier passed as source package `3.8.0`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r25-20260901_202457\ISSUE13_R25_CAUSE_ISOLATION_REPORT.md`.
|
||
- Pre-change snapshots: source `pre-issue13-r25-cause-isolation-excluding-pytest-cache-20260901_200300-0a7c670` (`826` files, mismatch zero), runtime diagnostic `pre-issue13-r25-runtime-diagnostic-20260901_202340` (`7` files, mismatch zero), and pre-Turbo-model inventory `pre-r25-turbo-model-add-20260901_201506`. Accepted completed source snapshot: `issue13-r25-cause-isolation-complete-excluding-pytest-cache-20260901_215430-0a7c670` (`833` files, mismatch zero). A5 remains HOLD; commit, push, tag, and release were not performed.
|
||
- After Queue `0/0`, stopped only dedicated 8190 backend PID `23804` and launcher PID `17664`; 8190 no longer listens. No 8188 process or frontend was stopped or restarted. Removed only the SHA-verified temporary runtime input copy; the original close-up image remains at `D:\SD_PIC\Neo\{AB283828-57FE-4C1C-A634-7F491B760453}.png`.
|
||
- The pinned `H3情報チェック` review accepted R2.5 as **cause isolation PASS / leading hypothesis C** with no mandatory correction, repeat GPU Gate, or Production change. Formal wording: `Issue #13 R2.5 complete / cause isolation PASS / leading hypothesis: continuation-count or cumulative feedback dependent / root cause unresolved / Motion LoRAs not necessary cause / Production unchanged / A5 HOLD`. Their possible long-run amplification role remains unresolved. Next priority is reporter exact workflow R3; only if it does not arrive, run one prompt-fixed `6x5` confirmation with the exact same prompt in all six chunks to separate continuation count from prompt progression.
|
||
|
||
## Advanced Phase A4 Chunk-scoped Refine GPU Experimental PASS (2026-09-01)
|
||
|
||
- Added the pure versioned `v3/refine_scope.py` resolver and appended `Refine Scope = All / Physical Group / Logical Chunk` plus a 1-based `Scope Index` to the existing public Experimental Selective Refine node. `All` is the default and retains the A3c execution path. Physical Group uses Assembly Plan order; Logical Chunk resolves to its owning physical group.
|
||
- Non-selected physical groups are skipped before prompt encode/refine-context adaptation, MODEL clone, seed derivation, and Sampling. Their original Video and Audio LATENT objects/tensors are returned unchanged. Selected seeds use the absolute physical-group index and therefore match All mode. Scope provenance remains status-only: no `refine_scope_contract` or other field was added to Assembly Plan, `second_pass_contract.version=1`, or Run Storage.
|
||
- Terminal Merge remains atomic. Selecting Logical Chunk 2 or 3 in `[2,3]` resolves to the same physical group, seed `12172123750489447764`, one Sampling pass, Video latent `f8847fb2...`, Audio latent `f28a2d05...`, decoded Video `50bf7d65...`, and decoded PCM `40b00232...`; Group 1 retains exact object/tensor/SHA identity.
|
||
- The dedicated 576x576 fixed-seed A4 GPU Matrix passed 5/5: G1 All Video+Audio 3x5 exact A3c G4 parity; G2 Physical Group 2 only; G3 Logical Chunk 2 only; G4 FL2VA Terminal Logical 2; and G5 the same Terminal Logical 3. G2/G3 also matched each other exactly. Every case produced 360 frames, exact 15 seconds, 24 fps, and 32 kHz stereo with finite, non-silent, non-collapsed, boundary-clean PCM and no OOM/NaN/Inf/decode/assembly error.
|
||
- API seconds were `470.776 / 144.220 / 138.590 / 480.440 / 250.880`; maximum VRAM was `15.566 GiB`, maximum backend Working Set `43.693 GiB`, and minimum available system RAM `2.157 GiB`. The 16GB GPU Gate passes with small headroom.
|
||
- Two initial evidence-only false STOPs were corrected without changing Production: cross-run Plan hashing included the existing volatile `_runtime_started_at`, and the first Terminal field-set check incorrectly used the non-Terminal G1 Plan as its reference. Revalidation requires identical field sets and all non-volatile values against the matching A3c non-Terminal/Terminal baseline; successful GPU Sampling was not repeated.
|
||
- The requested A3b/A3c saved-workflow fixture verifies that workflows lacking the two appended A4 widget values restore `Refine Scope = All` and `Scope Index = 1`. Focused A4/Second Pass regression passed `61`; post-GPU full pytest passed `968`; Manifest contains `234` entries; runtime verifier, Python syntax, JavaScript `7/7`, four source/runtime SHA pairs, and `git diff --check` pass.
|
||
- Snapshots: source `pre-v38-a4-chunk-scoped-refine-excluding-pytest-cache-20260901_174621-0a7c670` (`811` files, mismatch zero) and runtime `pre-v38-a4-gpu-runtime-w-excluding-pytest-cache-20260901_1845` (`393` files, mismatch zero). Only `v3/refine_scope.py`, `v3/second_pass.py`, `v3/second_pass_nodes.py`, and `v3/targeted_second_pass.py` were synchronized. Dedicated 8190 PID `35164` and launcher `15964` were stopped after Queue empty; user-owned 8188 PID `12636` and the frontend were untouched.
|
||
- Classification: **A4 Chunk-scoped Refine = GPU Experimental PASS**. Audio-targeted modes and the Selective node remain Production HOLD/Experimental under the existing upstream denoise-mask blocker policy. V3.5, Production Sampling, V3.8 Sampler, Frontend, workflows, Hi-Res Fix, Run Storage, package version, commit, push, tag, and release remain unchanged. A5 has not started. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a4-chunk-scoped-refine-gpu-matrix-20260901_185000\summary.json`.
|
||
- The pinned `H3情報チェック` final review formally accepted A4 as **GPU Experimental PASS** with no mandatory correction, repeat GPU generation, or additional Gate. It confirmed both evidence-only false STOPs were comparison-tool errors rather than A4 failures and required stopping before A5.
|
||
- Accepted source snapshot: `v38-a4-gpu-experimental-pass-accepted-excluding-pytest-cache-20260901_194500-0a7c670` (`819` files, mismatch zero; excludes `.git` and inaccessible `.pytest_cache`).
|
||
|
||
## Advanced Phase A3c Selective Refine GPU Matrix PASS (2026-09-01)
|
||
|
||
- **A3c = GPU Experimental PASS.** The public `H3 Continuum Selective Second Pass (Experimental)` node passed all five dedicated 576x576 fixed-seed GPU cases on ComfyUI_W/Core `0.34.2`: Video Only versus legacy V3.5 parity, Audio Only 1x5 seconds, Video + Audio 1x5 seconds, Video + Audio 3x5 seconds, and FL2VA Terminal Merge `[1] + [2,3]`.
|
||
- The stop-first G1 comparison was bit-exact: Video SHA-256 `07d7085276f34afe14a20c6eef27c13f85584755a664d8e8dbb856f54a1b4939`, Audio passthrough SHA-256 `f2302cbee59ec9a31a3a215fec2594269cb650dd269f6b6f136fc319b0f996f3`, refine seed `8643474934948678715`, effective SIGMAS hash, Sampling count, Audio object/tensor passthrough, and Assembly Plan all matched the unchanged legacy V3.5 node.
|
||
- Audio Only returned the exact First Pass Video LATENT SHA/object and changed only Audio. Video + Audio changed both target streams with preserved shapes and finite values. The 3x5 non-terminal case retained `[1]+[2]+[3]`; Terminal Merge retained two physical groups `[1]+[2,3]`, sampled the terminal pair once, and produced no partial pair.
|
||
- Every output was 24 fps and 32 kHz stereo at exactly 5 or 15 seconds. PCM was finite, non-silent, non-collapsed, duration-stable, and clean at structural boundaries. API time was `221.155 / 90.713 / 92.557 / 448.460 / 550.807` seconds; peak device use was `15.613 GiB`, peak backend Working Set was `44.08 GiB`, and minimum observed available system RAM was `2.594 GiB`. **16GB GPU: PASS, with low VRAM/RAM headroom under the A3c 3-chunk/Terminal workload.**
|
||
- Post-Gate validation passed `949` tests, Manifest `231/231`, compileall, JavaScript syntax `7/7`, runtime verifier, critical source/runtime SHA `6/6`, external evidence-tool syntax, and `git diff --check`. Accepted snapshots: source `v38-a3b-accepted-pre-a3c-excluding-pytest-cache-0a7c670-20260901_153752` (`807` files, mismatch zero) and runtime `pre-v38-a3c-runtime-w-excluding-pytest-cache-20260901_153758` (`393` files, mismatch zero).
|
||
- Classification remains separated: Video Only is a Production-compatible candidate; Audio Only and Video + Audio are GPU Experimental PASS / Production HOLD; the Selective Refine node remains Experimental. Upstream denoise-mask blocker policy is unchanged. Production Sampling, V3.5 node/schema, Frontend, workflows, Hi-Res Fix, Run Storage, package source version, commit, push, tag, and release were not changed. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a3c-selective-refine-gpu-matrix-20260901_154935\summary.json`.
|
||
- A3c used the approved minimal six-file runtime sync. ComfyUI_W's three package-label files (`version.py`, `metadata.ini`, `pyproject.toml`) retain the pre-existing runtime-copy label `3.7.0` and were intentionally not expanded into this GPU-only sync; the source remains `3.8.0`. **Known runtime packaging debt — not an A3c functional failure.** It did not affect registered V3.8/Selective nodes or Sampling, but must be resolved before public release or formal runtime-package finalization.
|
||
- The dedicated 8190 Queue was empty before shutdown; only its actual backend PID `15948` and venv launcher PID `36556` were stopped. Port 8190 has no listener. The user-owned 8188 backend remains listening as PID `12636` and was untouched.
|
||
- The pinned `H3情報チェック` review formally accepted A3c as GPU Experimental PASS with no mandatory code fix or repeat GPU run. Accepted source snapshot: `v38-a3c-gpu-experimental-pass-accepted-excluding-pytest-cache-20260901_162409-0a7c670` (`811` files, mismatch zero).
|
||
- Before A4, keep the upstream Production blocker and treat skipped physical groups as strict no-work/pass-through: no MODEL clone, conditioning, Sampling, seed consumption, or Video/Audio object replacement. Never refine only one half of Terminal `[2,3]`; selected-group seeds must match All mode; Assembly Plan and Run Storage contracts must remain unchanged.
|
||
|
||
## V3.8 Reliability Diagnostics CPU Gate PASS (2026-09-01)
|
||
|
||
- The existing `H3ContinuumSamplerV38` now appends read-only post-generation diagnostics without changing its accepted Aspect / Preset / Custom MP, Generation Mode / Review Action, socket, output, Frontend, or saved-workflow schema. V3.7 and earlier classes remain unchanged. `Report Detail = Off` skips all new reliability work; Basic stays silent for normal Frame/Audio/Core checks and Detailed Report emits per-physical-group evidence.
|
||
- Frame Alignment reports the existing Video Guide source-to-aligned frame decision and validates each physical Video latent against the H3 17k+5 grid. Audio validation inspects only `[B,32,2,T]` shape, metadata T, and `audio_grid_offset` for normal and Terminal Merge groups. Core compatibility reuses the existing feature-based advisory after successful generation and never gates or changes a route.
|
||
- `Visual conditioning load — peak per physical group` includes only resolved Reference Image areas plus Video Guide spatial/temporal ratios. It does not multiply by chunk count or Run Storage reuse, and it does not claim VRAM/time prediction. Still Image Guide is excluded from the numeric total and reported separately as active. Shared pure Reference Image and Video Guide resolvers keep the Production resize/frame contracts authoritative; diagnostics perform no resize, encode, clone, hash, pixel read, or device transfer.
|
||
- New reliability coverage adds `31` tests. Focused regression passed `167`; formal full validation passed `873`; Manifest `221/221`, Python syntax, runtime verifier (`3.8.0`, native PackedLayout, Fixed 3x5 planning), and source/runtime SHA-256 for the four changed Production files pass with only the known `pynvml` warning. Non-status output identity and Audio Tensor object/data-pointer/content identity are fixed by regression tests.
|
||
- Source snapshot: `pre-v38-reliability-diagnostics-approved-excluding-pytest-cache-20260901_004946-0a7c670` (`742` exact files, excluding only `.git` and ACL-inaccessible `.pytest_cache`). Runtime snapshot: `pre-v38-reliability-diagnostics-runtime-20260901_010833` (three pre-existing files exact; the new reliability module was absent). Only `reference.py`, `reference_video.py`, `v3/driving_nodes.py`, and `v3/reliability_v38.py` were synchronized to ComfyUI_W; no process was stopped or restarted. GPU generation, commit, push, tag, and release were not performed.
|
||
|
||
## V3.8 Easy Loader Bypass Enable Foreground PASS (2026-09-01)
|
||
|
||
- Image, Audio, and Video loaders now share one `web/easy_bypass_toggle.js` rendering contract. Native `node.mode = BYPASS` remains the source of truth and keeps the purple/transparent node appearance; after ComfyUI draws the bypassed widgets at `editorAlpha=0.2`, the shared helper redraws only `Enable Image`, `Enable Audio`, or `enable_video` at `editorAlpha=1`.
|
||
- The existing native widget remains in the normal widget list and retains ComfyUI's hit-test/callback path. It is visible and clickable while bypassed, and one click sets `node.mode` back to Always. Serialization, reload alignment, external Bypass/Unbypass, and bidirectional mode synchronization remain unchanged.
|
||
- Focused Image/Audio/Video regression passed `21`; full source passed `842`. Node syntax, Manifest `219/219`, `git diff --check`, and source/runtime SHA for the shared helper pass. Source snapshot: `pre-easy-loader-bypass-toggle-foreground-excluding-pytest-cache-20260901_000938-0a7c670` (`742` files, zero mismatch). Runtime target snapshot: `pre-easy-loader-bypass-toggle-foreground-runtime-20260901_001150` (one exact file).
|
||
- Only the shared frontend helper was synchronized to ComfyUI_W runtime. No backend or frontend process was stopped or restarted; a browser page reload is sufficient to load the updated module. Commit, push, tag, and GitHub Release remain unperformed.
|
||
|
||
## V3.8.0 Release Preparation Gate PASS (2026-08-31)
|
||
|
||
- **V3.8.0 Release Preparation Gate is PASS.** `version.py`, `pyproject.toml`, and `metadata.ini` now identify `3.8.0`; the English/Japanese README titles, package information, package validation text, CHANGELOG, and release-metadata assertions agree. Historical V3.7 entries, V3.7 node IDs, compatibility documentation, and the unchanged V3.7 Production path remain intact.
|
||
- README.md and README_JA.md now introduce Easy Mode, the Core+Continuum Strict Minimal workflow, the normal V3.8 workflow, shared aspect-aware Resolution presets, Review Each Chunk, Continue, Regenerate Current, Finish Remaining, Driving Audio, Easy Loader bypass behavior, and the partial-Review Second Pass limitation. Both public V3.8 workflow links resolve. The 16GB wording is conservative: Gate PASS with small headroom, Strict Minimal `14.96 GiB`, and complex accepted gates about `15.5-15.6 GiB`; no universal 16GB guarantee is claimed.
|
||
- `CHANGELOG.md` has a new `3.8.0` section with Added, Improved, and Notes. `PACKAGE_INFO_JA.txt` and `PACKAGE_VALIDATION.txt` record the two public workflows, Review/Smart lifecycle, Terminal atomicity, Driving Audio, current validation counts, small VRAM headroom, and the Second Pass limitation. Prompt / Context Builder is not presented as a V3.8.0 feature.
|
||
- The V3.8 Final Integration status remains PASS. G4 remains **PASS with prior transient non-reproducible parity event** and its diagnostics/evidence are retained; G5 Driving Audio remains PASS for the 20-second, 32 kHz stereo Review, Smart Regenerate, Finish Remaining, and restart sequence. User-facing README text does not promote the non-reproducible G4 event as a normal Known Issue.
|
||
- Release Preparation validation passed: full source `841 passed`; Manifest `219/219`; runtime verifier reported package `3.8.0`, native PackedLayout, and Fixed 3x5 prompt planning; Python `174/174`; JavaScript `7/7`; README relative links `41/41`; version consistency `9/9`; `git diff --check`; and final protected/non-authorized SHA `254/254` against the pre-Preparation snapshot. The first full pytest attempt was invalidated only by the inherited ACL-denied user Temp root; the accepted rerun used `D:\Codex\_tmp` and passed all 841 tests.
|
||
- Snapshot: `pre-v380-release-preparation-excluding-pytest-cache-20260831_233524-0a7c670` (`742` files, zero hash mismatch, excluding only `.git` and ACL-inaccessible `.pytest_cache`). Natural-Japanese lint completed in quick/tech mode; its 23 info findings were reviewed as technical-name, historical-section, or existing-style signals with no blocking issue.
|
||
- This is a prepared local release candidate only. No commit, push, tag, GitHub Release, workflow modification, Production implementation change, or backend operation was performed. Tag candidate is `v3.8.0`; publication requires separate explicit approval.
|
||
|
||
## V3.8 Workflow Cleanup Gate PASS (2026-08-31)
|
||
|
||
- **V3.8 Workflow Cleanup Gate is PASS.** `examples/workflows/MiniMax_H3_Continuum_Easy_V38.json` is now a 20-node Strict Minimal workflow using only ComfyUI Core plus H3 Continuum nodes. It removes the enabled rgthree Turbo LoRA, two `easy cleanGpuUsed` nodes, KJNodes Sage/ChunkFeedForward patches, PreviewAny utilities, and their links. The default graph is pure T2VA: all optional First/Last/Reference/Driving Audio loaders remain connected but native-bypassed, and Core `BasicScheduler` uses 20 steps after Turbo removal.
|
||
- Added the 19-node public `examples/workflows/MiniMax_H3_Continuum_V38.json`. It exposes the normal `H3ContinuumSamplerV38` with Aspect/Resolution, Full Run/Review Action, Run Storage/Regenerate, First/Last, three Reference Images, and Driving Audio while excluding Second Pass, Upscale, Spectrum, Turbo LoRA, rgthree, Easy-Use cleanup, and KJNodes. The 24 saved widget values exactly align with their named sequence; Generation Mode and Review Action remain the final two widgets.
|
||
- Workflow lint/save-reload coverage rejects rgthree, Turbo/Power LoRA, Easy-Use cleanup, Spectrum, Hi-Res/Second Pass/Upscale, MiniMaxChunkFeedForward, SageAttention patches, and any non-Core third-party `cnr_id`. It validates unique node/link IDs, both link directions, required connections, native bypass modes, Core titles, output chains, JSON roundtrip, and V3.8 widget positions.
|
||
- Dedicated 8190 Strict Minimal T2VA smoke passed without KJ workflow nodes: fixed seed, 1x10 seconds, Euler/simple 20 steps, Draft Landscape 736x416. Sampling/Decode/Assembly/Save completed in 310.35 seconds backend / 312.03 seconds API. Output is 240 frames, 24 fps, exact 10.000 seconds, 32 kHz stereo, finite audio, and non-black/non-collapsed video; no OOM or NaN/Inf occurred. Peak device VRAM was 14.96 GiB, peak process RSS 44.83 GiB, and minimum available RAM 6.30 GiB. The dedicated backend used the existing ComfyUI global `--use-sage-attention` option but no Sage/KJ workflow node.
|
||
- Source Gate passed `32` tests; final full source passed `841` tests. Manifest, runtime verifier, Python/JavaScript syntax, `git diff --check`, Critical source/runtime SHA, and all 172 protected Production/source files versus the pre-Cleanup snapshot pass. Snapshot: `pre-v38-workflow-cleanup-gate-excluding-pytest-cache-20260831_223800-0a7c670` (738 files, zero hash mismatch, excluding only `.git` and ACL-inaccessible `.pytest_cache`). Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-workflow-cleanup-gate-20260831_223800`.
|
||
- The first external smoke-runner attempt stopped before Queue submission because its API call omitted `client_id`; only the evidence runner was corrected. Dedicated 8190 was stopped after Queue empty; user-owned 8188 remained PID 35120 and was not operated. A Performance Optional workflow is not needed and was not created. Final Integration Gate rerun, README/CHANGELOG/version, commit, push, tag, and release were not performed.
|
||
|
||
## V3.8 Finish Remaining Phase F-G3 GPU Gate PASS (2026-08-31)
|
||
|
||
- **Phase F-G3 Finish Remaining GPU Gate is PASS.** A fresh fixed-seed T2VA 3x10-second Full baseline completed three chunks, and a separate fresh Review run generated Chunk 1 to `review_ready` before one `Finish Remaining` Queue exactly reused Chunk 1 and sampled only Chunks 2 and 3. The Finish Queue executed sampler node 305, was not cache-skipped, performed exactly two physical Sampling groups, and finalized all three chunks as `complete`.
|
||
- Full versus Finish parity is exact for every logical chunk's seed, Prompt hash, video/audio latent tensor SHA-256, and assembly plan, plus decoded RGB24 video and float32 32 kHz stereo PCM. Final media is 720 frames, 736x416, 24 fps, and 30.000 seconds. The final Review head is revision `4762ea4bd196c3f3`, branch boundary 0, effective nonce 0, review unit 3, and three validated chunks.
|
||
- Sampling/API seconds were Full 252+324+325 / 1035.089, Review Queue 1 247 / 293.268, and Finish Remaining 321+319 / 759.298. Peak device VRAM was 15.57 GiB; maximum process RSS was 48.53 GiB and minimum available system RAM was 4.18 GiB. The 16GB Gate classification remains PASS with small VRAM headroom.
|
||
- The API prompt retained `Finish Remaining`; the separate frontend one-shot regression passed `7/7`, covering reset to `Continue / Next` after accepted Queue. Post-Gate validation passed: full source `835 passed`, Manifest `217/217`, runtime verifier, Python/JavaScript syntax, eight source/runtime SHA-256 pairs, and `git diff --check`.
|
||
- The first diagnostic attempt stopped after a valid Review Queue 1 because the external evidence runner compared whole safetensors container SHA across separate Full/Review revisions; all required Sampling parity fields were exact. Only the external runner was corrected, and the complete Gate was rerun under fresh `retry1` Run Storage names. Production code was not changed.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g3-20260831_130650\PHASE_F_G3_PASS_REPORT.md` and `gpu_retry1/g3_summary.json`. Snapshot `pre-v38-review-phase-f-g3-finish-gpu-gate-20260831_130526-0a7c670` contains 724 exact source files excluding `.git`/`.pytest_cache`.
|
||
- The dedicated 8190 backend was stopped after Queue empty; port 8188 remained HTTP 200 and untouched. Optional Smart-branch Finish Remaining, G4, G5, workflow/version/release text, commit, push, tag, and release were not performed.
|
||
|
||
## V3.8 Smart Regenerate Phase F-G2 GPU Gate PASS (2026-08-31)
|
||
|
||
- **Phase F-G2 Smart Regenerate GPU Gate is PASS.** A fresh fixed-seed T2VA 3x10-second Review run generated Chunk 1 then Chunk 2, followed by three `Regenerate Current` queues. The Smart target remained Chunk 2, effective nonce advanced `0 -> 1 -> 2 -> 3`, all four derived seeds and all four Chunk 2 video latent SHA-256 values were unique, and Chunk 1 seed/prompt/video/audio identity remained exact with zero Regenerate Sampling.
|
||
- Sampler node 305 executed and was absent from `execution_cached` for Initial Q1/Q2, R1/R2/R3, and Continue. Each Queue sampled exactly one new/replaced physical group; Run Storage provided all prefix reuse. Every stored video/audio latent was finite, with no OOM, NaN/Inf, Decode error, or Assembly error.
|
||
- Canonical review head advanced in order through revisions `4762ea4bd196c3f3` (nonce 0), `a2b5a5c96b64bf3f` (nonce 1), `61c28a5b4322b6eb` (nonce 2), and `bd82c8661d8863ab` (nonce 3). After stopping and restarting only the dedicated 8190 backend, Production `RunStorageController.find_latest_review_head()` restored revision `bd82c8661d8863ab`, prefix 2, `review_unit=2`, branch 2, and nonce 3 from disk.
|
||
- Restarted Continue exactly reused Chunk 1 and nonce 3 Chunk 2, generated only Chunk 3, retained revision `bd82c8661d8863ab` / branch 2 / nonce 3, and finalized three chunks as `complete`. Existing frontend one-shot reset behavior remains covered by the Phase E UI regression; the GPU Gate used explicit API actions.
|
||
- Sampling/API seconds were Q1 254/311.379, Q2 323/404.905, R1 323/404.718, R2 324/406.180, R3 324/404.887, and restarted Continue 327/457.637. Peak device VRAM was 15.59 GiB; maximum process RSS was 48.24 GiB and minimum available system RAM was 3.16 GiB. The 16GB classification remains PASS with small VRAM headroom.
|
||
- Post-Gate validation: full source `835 passed`, Manifest `217/217`, runtime verifier, 57 Python files, 7 JavaScript files, eight source/runtime SHA-256 pairs, and `git diff --check` all PASS. Only the known `pynvml` and inherited `.pytest_cache` ACL warnings remain.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g2-20260831_113611\PHASE_F_G2_PASS_REPORT.md` and `gpu_retry1/g2_summary.json`. Snapshot `pre-v38-review-phase-f-g2-gpu-gate-20260831_113611-0a7c670` captured 724 source files with zero missing, extra, or hash mismatch, excluding `.git` and inherited `.pytest_cache`.
|
||
- The first diagnostic attempt stopped after a successful Production R1 because the evidence helper ignored Smart prefix `storage_revision_id`; the external evidence runner alone was corrected and the complete Gate was rerun under a fresh run name. Production code was not changed. The dedicated 8190 backend was stopped after Queue empty; port 8188 was unreachable before/after and untouched. G3, workflow/version/release text, commit, push, tag, and release were not performed.
|
||
|
||
## V3.8 Review Each Chunk Phase F-R1 Recovery PASS (2026-08-31)
|
||
|
||
- **The Phase F-R1 cache recovery is PASS.** `H3ContinuumSamplerV38.IS_CHANGED()` now returns stable `False` for Full Run and `NaN` only for Review Each Chunk. V3.7, Easy, Production Sampling, Run Storage schema/identity, seed derivation, prompt identity, and variation semantics are unchanged. There is no frontend token, queue monkey patch, or manifest fingerprint.
|
||
- The actual ComfyUI Core cache harness proves Full Run signatures remain stable while Review Continue, Regenerate Current, and Finish Remaining change every Queue. Focused cache tests passed `8`, related V3.7/V3.8/Easy/Terminal Merge/Run Storage tests passed `259`, and the full source suite passed `835`; the runtime verifier, Python/JavaScript syntax, all `217` Manifest entries, and `git diff --check` passed with only the known `pynvml` and inherited `.pytest_cache` ACL warnings.
|
||
- Recovery resumed the stopped G1 run at one stored chunk / `review_ready` and resubmitted the failed Queue 2 API prompt unchanged. Queue 2 executed sampler node 305, preserved Chunk 1 exactly, generated only Chunk 2, updated the manifest to two chunks / `review_ready`, and completed in 430.143s. Queue 3 submitted the same prompt again on the same backend, did not cache sampler 305, preserved Chunks 1-2 exactly, generated only Chunk 3, and completed at three chunks / `complete` in 450.204s.
|
||
- Final G1 exactly matches the existing G0 Full baseline for every chunk's seed, prompt hash, plan, video latent, and audio latent, plus decoded RGB24 video, PCM float32 32 kHz stereo audio, frame count, geometry, fps, channels, and duration. Queue 2 / Queue 3 peak device VRAM was 15.43 / 15.49 GiB, so the existing 16GB Gate classification remains PASS with small headroom.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-r1-20260831_111111\gpu-recovery\recovery_summary.json`. Accepted source snapshot: `pre-v38-review-phase-f-r1-cache-invalidation-approved-excluding-pytest-cache-20260831_110319-0a7c670` (719 files, exact path/hash match). Accepted runtime snapshot: `pre-v38-review-phase-f-r1-runtime-sync-excluding-pytest-cache-20260831_110749-runtime` (383 files, exact path/hash match).
|
||
- Only `v3/driving_nodes.py` was synchronized to ComfyUI_W with exact source/runtime SHA-256 equality. The dedicated 8190 backend was stopped after an empty Queue. Port 8188 was already unreachable before this Recovery run and remained untouched. Smart Regenerate G2 and later Phase F cases were not executed automatically; no workflow/version/release text, commit, push, tag, or release was performed.
|
||
|
||
## V3.8 Review Each Chunk / Smart Regenerate Phase F GPU Gate STOP (2026-08-31)
|
||
|
||
- **Phase F is FAIL / STOP; Review Each Chunk and Smart Regenerate are both HOLD.** The fixed-seed T2VA 3x10-second, 20-step G0 Full baseline completed with three chunks, 720 frames / 30.000s / 24 fps / 32 kHz stereo, no OOM or NaN/Inf, and exact latent/decoded evidence. G1 Review Queue 1 correctly generated only Chunk 1, saved `review_ready`, and exactly matched G0 Chunk 1 for seed, prompt hash, video/audio latent SHA-256, and plan.
|
||
- G1 Review Queue 2 submitted the exact same V3.8 sampler inputs as Queue 1 (canonical sampler-input SHA-256 `dbc177fd9c7a25a8da6b328886d0fcd1bc024157d16bba15b1109f54970510ee`). ComfyUI history reported `execution_cached` for `H3ContinuumSamplerV38` node 305 and its Decode/Assembly path. Queue 2 finished in 6.075s with zero Sampling, one stored chunk, unchanged `review_ready` manifest mtime, and the same partial decoded output. Expected Chunk 1 reuse plus Chunk 2 generation did not occur.
|
||
- Phase F rules prohibited a test-only cache nonce or relaxed acceptance. G2 Smart nonce 0->1->2->3, restart, Finish Remaining, FL2VA Terminal Merge, and Driving Audio were not run after this first Production failure. Repeated identical `Regenerate Current` queues are subject to the same missing Queue-to-cache invalidation contract, so Smart Regenerate cannot be promoted on the Phase D unit evidence alone.
|
||
- GPU metrics: G0 API 1053.884s, peak device VRAM 15.472 GiB, peak process RSS 46.555 GiB, minimum system available RAM 0.575 GiB; G1 Queue 1 API 297.318s, peak device VRAM 15.527 GiB. This is a 16GB Gate pass with small VRAM and system-RAM headroom, not an "余裕あり" result.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-gpu-gate-20260831_101000\PHASE_F_STOP_REPORT.md` and `gpu-g0-g1/gate_summary.json`. Added only the diagnostic runner `tools/v38_review_phase_f_gpu_gate_runner.py`; Production code was not changed. A later phase must design a V3.8 Review-specific cache invalidation / Queue intent token contract without changing seed, prompt, sampling identity, or variation semantics, add fixed-seed Continue and repeated-Regenerate integration coverage, then rerun Phase F from G0.
|
||
- Runtime verifier and Review UI harness 7/7 passed. Codex Chrome could not complete live port-8190 UI inspection because its browser extension returned `ERR_BLOCKED_BY_CLIENT`; this is recorded as unverified rather than a product UI failure. The dedicated 8190 backend was stopped after an empty Queue; the user-owned 8188 Stability Matrix backend remained HTTP 200 and was not stopped or restarted. No workflow/version/release text, commit, push, tag, or release was performed.
|
||
|
||
## V3.8 Review Each Chunk / Smart Regenerate Phase E (2026-08-31)
|
||
|
||
- **Phase E Review UI / Queue Intent / One-shot Action source Gate PASS.** Only `H3 Continuum Sampler V3.8` exposes `Generation Mode` (`Full Run` default / `Review Each Chunk`) and `Review Action` (`Continue / Next` default / `Regenerate Current` / `Finish Remaining`). The two controls are appended after every pre-Phase-E V3.8 widget so existing positional widget values keep their indices. V3.7 and Easy schemas remain unchanged.
|
||
- V3.8 forwards only `generation_mode` and `review_action`; Phase B-D remain the backend source of truth for canonical head, target unit, branch boundary, effective nonce, Terminal Merge, validated prefix, and physical-group cap. Full Run normalizes its submitted frontend action to Continue and the backend pure resolver independently ignores stale actions. Smart plus Manual Regenerate remains a backend validation error and the Review frontend does not calculate or rewrite its target/nonce.
|
||
- Frontend audit used the installed ComfyUI frontend package 1.49.6 source map. The official widget lifecycle runs `beforeQueued`, serializes with `graphToPrompt`, receives a successful `/prompt` response, then invokes widget `afterQueued`; failed submission does not invoke `afterQueued`. Phase E reuses the existing `beforeQueuePrompt` hook and attaches only the formal widget `afterQueued` callback—no `app.queuePrompt`/`api.queuePrompt` monkey patch or competing queue hook.
|
||
- On Full -> Review selection, Run Storage changes from Off to Save + Auto Resume; Review -> Full preserves it. Review Action is hidden in Full without removing/reordering widgets. Regenerate/Finish are reset to Continue only after accepted Queue, while the already-submitted prompt retains the original action; failed Queue preserves the action. Workflow load normalizes saved one-shot actions to Continue and never auto-queues.
|
||
- Backend Status now reports Review ready progress, atomic Terminal Merge ranges, Smart Regenerate/preserved prefix (variation only in Detailed Report), Finish Remaining totals, and sequence completion from Phase D manifest/controller facts. Incomplete Review plus refine-context capture adds the explicit warning that partial-review Second Pass refinement is not yet supported; completed-sequence behavior is unchanged.
|
||
- Validation: Phase E plus B-D focused `86 passed`; V3.7/V3.8/Easy/Reference/Driving Audio/Terminal Merge/Run Storage/frontend related `362 passed`; full source `827 passed`. Python/JavaScript syntax, `git diff --check`, and all 214 Manifest entries pass. Only the known `pynvml` warning and inherited ACL-inaccessible repository `.pytest_cache` warning remain; tests use a writable `D:\Codex\_pytest_temp` basetemp.
|
||
- Accepted snapshot: `pre-v38-review-phase-e-ui-approved-excluding-pytest-cache-20260831_025634-0a7c670` contains 715 files with zero missing, extra, or SHA-256 mismatch. The standard snapshot attempt stopped on the inherited ACL-inaccessible `.pytest_cache` before implementation.
|
||
- Runtime sync, live UI, backend/frontend restart, GPU execution, workflow JSON, package/schema/contract version, README/CHANGELOG release text, commit, push, tag, release, and Phase F were not performed.
|
||
|
||
## V3.8 Review Each Chunk / Smart Regenerate Phase D (2026-08-31)
|
||
|
||
- **Phase D Review Run Storage Lifecycle / Smart Regenerate Branch Persistence PASS.** Run Storage now distinguishes `in_progress`, intentional `review_ready`, abnormal `interrupted`, and configured-sequence `complete`. `review_ready` is written only after an atomic physical group has sampled, validated, updated state, committed every logical entry, passed stored-file/prefix validation, and reached successful Production finalization.
|
||
- Review manifests add only non-identity metadata: `review_control_version: 1`, `review_unit {start,end,physical_group}`, `review_pause_reason: review_each_chunk`, `branch_regenerate_from`, and `effective_reroll_nonce`. Run Storage schema remains 2 and Sampling Contract remains 5; Generation Mode, Review Action, pause state, and branch metadata are not part of sampling identity or seed derivation.
|
||
- Canonical state selection reuses the existing compatible `nonce_lineage_sha256` and deterministic branch ordering: highest effective nonce, then persisted `updated_utc`, then revision ID. Candidates are revalidated for contract/revision SHA, global sampling/prompt/reference/Driving Audio lineage, chunk contract, contiguous prefix, safetensors size/SHA, status, review-unit geometry, branch boundary, and nonce. A corrupt newest branch falls back to the latest valid compatible head; filesystem enumeration order and mtime are not used.
|
||
- Production keeps Phase B policy pure. Run Storage supplies validated disk facts to `resolve_review_execution()`, then applies its boundary, effective nonce, and Phase C physical-group cap internally. Review controls remain private Python arguments and are absent from every public node schema pending Phase E.
|
||
- Consecutive Smart Regenerate of the same Chunk 2 persists nonce `0 -> 1 -> 2 -> 3`, produces distinct existing-contract chunk seeds, and keeps the current branch on Continue/Finish. `in_progress` or `interrupted` resume retains the same boundary and nonce. Review-to-Full reuses the compatible branch without making review policy identity; Full Run-derived complete manifests receive no `review_unit`.
|
||
- Terminal Merge remains atomic: the final pair is stored as one `review_unit` spanning both logical chunks, Smart Regenerate targets its first chunk, and a manifest that names only one terminal half is rejected. Finish Remaining records the actual final generated physical group rather than a stale earlier review unit.
|
||
- Restart coverage writes manifests and safetensors, discards the controller, reloads from disk, and restores Continue, Regenerate Current, Finish Remaining, and Full Run branch decisions. Legacy revisions without review metadata remain valid for normal resume and are never accepted as Smart Regenerate Current heads.
|
||
- Validation: Phase D focused `25 passed`, Phase C combined focused `43 passed`, related Run Storage/Terminal Merge/Driving Audio/Reference/V3.7/V3.8 `269 passed`, and full source `820 passed`. Python syntax, `git diff --check`, and all 213 Manifest entries pass with only the known `pynvml` warning.
|
||
- Accepted snapshot: `pre-v38-review-phase-d-lifecycle-approved-excluding-pytest-cache-20260831_021636-0a7c670` contains 711 files with zero missing, extra, or SHA-256 mismatch. The standard snapshot attempt stopped on the inherited ACL-inaccessible `.pytest_cache` before implementation.
|
||
- Phase E remains required to expose Generation Mode / Review Action only on the normal V3.8 sampler, implement one-shot reset and user-facing Status, and connect frontend queue intent. Frontend, workflow JSON, GPU, runtime sync/restart, package/schema versions, README/CHANGELOG release text, commit, push, tag, and release were not changed in Phase D.
|
||
|
||
## V3.8 Review Each Chunk / Smart Regenerate Phase C (2026-08-31)
|
||
|
||
- **Phase C Physical Review Execution Cap PASS.** Production-internal `max_new_physical_groups: int | None = None` is implemented in `v2/sequence.py`. `None` preserves full execution; a positive limit counts only newly sampled physical groups after the reusable prefix. The option is not public schema and is not part of sampling identity, Run Storage settings, or seed derivation.
|
||
- Normal chunks count as one physical group after sampling, validation, entry/state update, optional Run Storage commit, and cleanup. The final two logical FL2VA chunks count as one atomic Terminal Merge physical group; both logical entries are validated and committed before the group is counted or execution may stop. An inherited half-pair session is trimmed to the valid prefix and the complete terminal pair is regenerated, so a half pair is never returned.
|
||
- Limited execution may return only a non-empty contiguous prefix. `session`, state, entries, video/audio latents, and `assembly_plan` contain only completed chunks; no future dummy entries are created. Full execution retains the configured-chunk completeness check. The latent report adds one concise physical-review progress line.
|
||
- V3 nodes now distinguish configured chunks from completed entries. Terminal Merge detection still uses the configured sequence, and `preserve_final_frame` is false for an incomplete prefix. In 3x5-second First+Last coverage, queue-equivalent call 1 returns only chunk 1 without Last Frame; call 2 reuses chunk 1 and produces terminal chunks 2-3 atomically with Last Frame applied.
|
||
- Full-versus-incremental mock parity passes for normal and Terminal Merge sequences across logical entries, seed, prompt hash, video/audio latents, physical decode grouping, and assembly plan. Partial `capture_refine_context` remains explicitly incomplete; Second Pass support was not added.
|
||
- Validation: Phase C focused `17 passed`, related Phase B/Terminal Merge/Run Storage/V3.7/V3.8 `216 passed`, full source `794 passed`, changed-file Python syntax, and `git diff --check` PASS. Only the known `pynvml` warning remains. Protected `run_storage.py`, `v3/review_control.py`, `web/project_id.js`, `MANIFEST.sha256`, `v3/driving_nodes.py`, and `v3/plan.py` match the Phase C snapshot by SHA-256.
|
||
- Accepted snapshot: `pre-v38-review-phase-c-physical-cap-approved-excluding-pytest-cache-20260831_015840-0a7c670` contains 707 files with zero missing, extra, or SHA-256 mismatch. The standard snapshot attempt stopped on the inherited ACL-inaccessible `.pytest_cache` before implementation.
|
||
- Phase D remains required for persisted `review_ready` lifecycle, canonical review-head selection, review metadata, pause reason, branch boundary/effective nonce inheritance, and Smart Regenerate nonce advancement. No frontend/UI, GPU execution, workflow, schema/contract version, commit, push, or release work was performed in Phase C.
|
||
|
||
## V3.8 Review Each Chunk / Smart Regenerate Phase B (2026-08-31)
|
||
|
||
- **Phase B pure resolver PASS.** `v3/review_control.py` defines the V3.8 Generation Mode, Review Action, atomic `ReviewUnit`, non-identity review-pause metadata, and a side-effect-free `resolve_review_execution()` contract. It imports no ComfyUI runtime, torch, filesystem, Run Storage, or sampling code and is not connected to Production execution yet.
|
||
- Full Run resolves to no physical-group cap. Review Continue resolves at most one next physical group, including the final two logical chunks as one Terminal Merge unit. A prefix that ends inside the Terminal Merge pair is rejected as an impossible internal state.
|
||
- Smart Regenerate requires a `review_ready` head, or a `complete` head that explicitly contains `review_unit` metadata. It rejects an empty/missing head and Smart plus Manual Regenerate ambiguity. A completed Full Run without review metadata is not guessed as "Current" and must use the existing Manual Regenerate path.
|
||
- Branch continuation is explicit: the resolver adds `latest_branch_regenerate_from` to the minimum audit inputs because effective nonce alone cannot reconstruct the branch boundary. Continue/Finish/Full after a reviewed regeneration inherit both the branch boundary and effective nonce; consecutive Smart Regenerate requests advance the current branch nonce `1 -> 2 -> 3` without defining a new seed algorithm.
|
||
- `partial_review` describes the projected output prefix for later explicit Second Pass validation. V3.8 Review v1 does not claim incomplete-prefix `refine_context` / Second Pass parity and must not silently present prompt-only fallback as supported before its separate Extra High Gate.
|
||
- Validation is focused `36 passed`, related V3.7/V3.8/Run Storage/Terminal Merge `152 passed`, and full source `777 passed`; changed-file Python syntax passes with only the known `pynvml` warning. Production Sampling, `run_storage.py`, `v3/nodes.py`, frontend, GPU workflows, package/schema/contract versions, and `MANIFEST.sha256` are unchanged in Phase B.
|
||
- Accepted snapshot: `pre-v38-review-phase-b-pure-resolver-approved-excluding-pytest-cache-20260831_012741-0a7c670` contains 701 files with zero missing, extra, or SHA-256 mismatch. The standard snapshot attempt stopped on the inherited ACL-inaccessible `.pytest_cache` before implementation.
|
||
|
||
## H3 Continuum Load Video Gate (2026-08-31)
|
||
|
||
- **Gate PASS.** `H3ContinuumLoadVideo` / `H3 Continuum Load Video` is a minimal helper under `MiniMax H3/Continuum/Easy`, with exactly `Enable Video`, video upload/select, and `Force Rate` inputs plus standard `IMAGE` / `AUDIO` outputs. It is usable outside Easy workflows and does not create a Utility category.
|
||
- Decode is owned by the current ComfyUI Core `InputImpl.VideoFromFile(...).get_components()`. Continuum copies no VHS/PyAV/FFmpeg decode path. Force Rate `0` preserves the source batch; positive rates use Fraction-based previous-frame selection to drop/duplicate IMAGE frames while preserving nominal duration. Audio is returned unchanged and is never stretched or resampled.
|
||
- Enable OFF returns silent `ExecutionBlocker(None)` values for both branches before path resolution, validation, fingerprint stat, or decode. A video without audio returns IMAGE normally and blocks only AUDIO. ON validates with Core `exists_annotated_filepath`; the cache fingerprint includes enable state, file name, Force Rate, and modification time.
|
||
- `web/easy_load_video.js` binds the existing serialized Enable Video widget to ComfyUI native node modes: false -> Bypass mode 4 with the active theme's standard purple appearance; true -> Always mode 0. External Bypass and workflow reload synchronize back to the widget. No color value is hardcoded, no widget is inserted/removed, and backend ExecutionBlocker behavior remains available for API execution without the frontend extension.
|
||
- Validation is focused Easy loaders `20 passed`, actual-Core focused `10 passed`, and full source `741 passed`; all 209 Manifest entries, Python/JavaScript syntax, runtime verifier, live placement schema, source/runtime hashes, and `git diff --check` pass with only the known `pynvml` warning. Real Core decode of the existing G1 Gate MP4 returned IMAGE `(120,416,736,3)` and standard 32 kHz stereo AUDIO `(1,2,160768)` at Force Rate 24.
|
||
- Dedicated live registration on `127.0.0.6:8190` reported required inputs `enable_video,file,force_rate` and outputs `IMAGE,AUDIO`; its launcher/backend PIDs were stopped after the check. The user Stability Matrix instance at `127.0.0.6:8188` remained alive and was not restarted.
|
||
- Accepted snapshots: source `pre-h3-continuum-load-video-approved-excluding-pytest-cache-20260831_002937-0a7c670` (694 files, missing/extra/hash mismatch zero) and runtime `pre-h3-continuum-load-video-runtime-sync-20260831_004000-runtime` (one pre-existing file; the new loader did not exist). V3.7 Production, workflows, README, CHANGELOG, package version, GPU Sampling, commit, push, tag, and release are unchanged.
|
||
- Native-Bypass appearance snapshots: source `pre-h3-load-video-native-bypass-color-approved-excluding-pytest-cache-20260831_004918-0a7c670` (700 files, missing/extra/hash mismatch zero) and runtime `pre-h3-load-video-native-bypass-color-runtime-sync-20260831_005400-runtime` (one pre-existing shared helper; the new video frontend file did not exist).
|
||
- The still-running user backend at `127.0.0.6:8188` now lists `/extensions/ComfyUI-H3-Continuum/easy_load_video.js`; no backend restart is required for this frontend-only addition, but the ComfyUI page must be reloaded to import it.
|
||
|
||
## V3.8 Main Sampler Resolution Preset prototype (2026-08-31)
|
||
|
||
- **The user formally accepted the V3.8 Main Resolution Preset Gate as PASS.** This remains a tested sampler-facade candidate over unchanged V3.7 Production, not a frozen 3.8.0 specification, package-version change, release candidate, or workflow migration. `H3ContinuumSamplerV37` retains its existing `width` / `height` schema and saved-workflow contract.
|
||
- `v3/resolution.py` is now the single pure owner of the approved Easy/Main resolution contract: Auto from First Image with 16:9 fallback, Landscape 16:9, Portrait 9:16, Square 1:1, Draft 0.30 MP, Balanced 0.60 MP, Native 768 with 1344 long-edge priority, Custom 0.10-16.00 MP, 2.50 MP strong-warning threshold, and nearest 32-pixel alignment. Easy keeps compatibility aliases and resolves through the same helper.
|
||
- `H3ContinuumSamplerV38` / `H3 Continuum Sampler V3.8` replaces only its public `width` / `height` widgets with `Aspect`, `Resolution Preset`, and `Custom MP`. Before sampling it resolves geometry and passes only `width` / `height` into `H3ContinuumSamplerV37.run()`. Preset names do not enter Run Storage identity; existing geometry compatibility, refine context, Second Pass source geometry, continuation, conditioning, audio, and assembly paths remain owned by V3.7 Production.
|
||
- Existing `web/project_id.js` treats V3.8 as a Production sampler for project identity, Run Storage normalization, Regenerate From, diagnostics/debug/preview settings, reference-audio labels, and queue normalization. The only new UI behavior is native widget visibility for `Custom MP`; no dedicated V3.8 JS, dynamic sockets, widget splicing, or accordion was added.
|
||
- CPU acceptance is focused `97 passed`, frontend retry `40 passed`, and full source `729 passed`; all 206 Manifest entries, Python/JavaScript syntax, runtime verifier, and `git diff --check` pass with only the known `pynvml` warning. The V3.7 schema-invariance, all Aspect x Preset Easy/Main parity, Auto image geometry/fallback, 32-pixel alignment, Custom validation, resolved Run Storage identity, registration, and frontend parity are fixed by `tests/test_v38_resolution_preset.py`.
|
||
- GPU mini Gate is PASS 3/3 on RTX 5060 Ti 16 GiB using one 5-second chunk, fixed seed `927966463230733`, Sage, 20 evaluations, external Core VAE decode, and V3.5 assembly: Draft Landscape `736x416` at 175.824s; Balanced Portrait `576x1024` at 294.478s; Native 768 Auto with a 1024x1344 First Image `768x1024` at 484.487s. Every output passed 120 frames, 24 fps, exact 5.000s duration, 32 kHz stereo, finite audio, and non-black/non-collapsed sampled video checks. Peak device-wide VRAM was 15.33 GiB.
|
||
- Evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-main-resolution-gpu-gate-20260830_234000\gpu`. The five runtime implementation files match source by SHA-256. The dedicated `127.0.0.6:8190` Gate backend was stopped with zero remaining processes; the Stability Matrix `127.0.0.6:8188` instance remained alive and was not restarted.
|
||
- Accepted snapshots: source `pre-v38-main-resolution-preset-approved-excluding-pytest-cache-20260830_232727-0a7c670` (685 files, missing/extra zero) and runtime `pre-v38-main-resolution-runtime-sync-20260830_233637-runtime` (four pre-existing files; shared resolver did not previously exist). Workflow JSON, README, CHANGELOG, package version, commit, push, tag, and release are unchanged.
|
||
|
||
## Easy workflow prototype / Easy Load Audio + Driving Audio (2026-08-30)
|
||
|
||
- **This remains an Easy-workflow feasibility and specification prototype over V3.7 Production; it is not a frozen V3.8 feature set, final workflow, package-version decision, or release candidate.** Production Sampling and every V3.7 public node remain unchanged.
|
||
- `H3EasyLoadAudio` / `H3 Easy Load Audio` subclasses the current ComfyUI Core `comfy_extras.nodes_audio.LoadAudio`. It changes only schema identity/category/description and inherits Core loading, upload/player, validation, fingerprint, and file handling. It outputs AUDIO and owns no VAE.
|
||
- `web/easy_bypass_toggle.js` is the shared frontend owner for `Enable Image` and `Enable Audio`. `node.mode` is the single source of truth; ON uses mode 0, OFF uses native Bypass mode 4, external Bypass changes synchronize back, and the non-serialized toggle is removed temporarily during configure/serialize so Core filename/player widget positions remain intact. There is no polling and no backend Boolean/None-return path.
|
||
- `H3ContinuumEasyV38` now exposes optional `driving_audio` (display `Driving Audio`) and `audio_vae` (display `Driving Audio VAE`) and passes both directly to unchanged `H3ContinuumSamplerV37.run()`. Audio without a VAE retains the existing V3.7 validation; a VAE without Audio is ignored by the existing Production path and normal generated audio remains active.
|
||
- Native Bypass uses current ComfyUI prompt semantics: the bypassed loader is omitted and its downstream optional input becomes unconnected. GPU G0/G3 OFF confirmed that leaving Audio VAE connected while Driving Audio is absent returns to normal generated audio. Fast Groups Bypasser is not used on this route.
|
||
- GPU Gate is PASS 6/6 on ComfyUI_W / RTX 5060 Ti 16 GiB: G0 10s OFF, G1 10s ON, G2 20s/2-chunk ON, and G3 ON -> OFF -> ON. All outputs passed 480x640 geometry, 24 fps, exact 240/480-frame duration, 32 kHz stereo, finite audio, and no execution/OOM failure. The 20-second boundary delta was `0.114565`, below the observed `0.999` adjacent-delta quantile `0.530163`; no boundary click spike was detected. The three identical 10-second ON runs produced the same saved PCM SHA-256. Input-prefix versus saved-FLAC PCM was not bit-exact, so no new file-level bit-exact requirement was inferred beyond the existing in-memory V3.7 contract.
|
||
- Gate elapsed times were 189.089s / 171.138s / 348.234s / 167.914s / 160.616s / 167.037s. Peak device-wide VRAM was 15.05 GiB, peak backend RSS 46.02 GiB, and minimum available system RAM 5.05 GiB. Evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-easy-driving-audio-gpu-gate-20260830_223000\gpu-2`.
|
||
- `examples/workflows/MiniMax_H3_Continuum_Easy_V38.json` is now a 27-node candidate with one default-bypassed `H3 Easy Load Audio`, direct Driving Audio connection, and the existing Core Audio VAE shared with `VAEDecodeAudio`. Loader nodes retain their standard node display names; role labels are expressed by workflow groups, and Core node titles are never customized.
|
||
- CPU validation is focused `64 passed` before GPU, workflow-focused `57 passed`, and final full source `694 passed`; 203/203 Manifest entries, Python/JavaScript syntax, `git diff --check`, runtime verifier, live registration, and empty Queue pass with only the known `pynvml` warning. Accepted snapshots: source `pre-easy-load-audio-driving-audio-approved-excluding-pytest-cache-20260830_220541-0a7c670` (676 files, zero missing/extra/hash mismatch) and ComfyUI_W runtime `pre-easy-load-audio-driving-audio-runtime-w-approved-excluding-pytest-cache-20260830_221217-runtime` (372 files, zero hash mismatch).
|
||
|
||
## V3.8 Easy Mode v1.1 / Phase 3 GPU acceptance (2026-08-30)
|
||
|
||
- **Current status is a Phase 3 candidate, not a frozen or final V3.8 specification.** The GPU PASS establishes evidence for the tested Easy/Reference design only. Feature scope, UI, workflow composition, documentation, version 3.8.0, and release readiness remain open until the user explicitly approves a later decision.
|
||
- **Phase 3 GPU Gate is PASS (7/7)** on ComfyUI_W / NVIDIA GeForce RTX 5060 Ti 16 GiB. Every measured run used the same `minimax_h3_fl2va_pruned_int8_convrot.safetensors` model, non-LightX 4-step Turbo LoRA at strength 1.0, Euler/simple with 8 evaluations, Draft 0.30 MP, fixed base seed `927966463230733`, Run Storage Off, external Core VAE Decode, and V3.5 Exact Duration Assembly.
|
||
- The mandatory identical-Prompt/Seed matrix passed for Sage, Sage+Sol, Sage+Spectrum, and Sage+Sol+Spectrum using I2VA + Reference 1 at 10 seconds. API elapsed times were 302.979s, 211.546s, 146.037s, and 142.920s respectively. These are correctness observations from one warm runtime, not a standalone performance benchmark.
|
||
- Functional Sage cases also passed: First+Last+Reference 15s = 2x7.5s at 272.906s; Reference Image 2 only 20s = 2x10s at 371.483s; no-Reference T2VA 30s = 3x10s at 514.388s. The gap case compacted to one active reference. Hybrid Picture numbering used `<Picture 2>` for First+Reference and `<Picture 3>` for First+Last+Reference without H3C-P102/P103 warnings.
|
||
- All seven MP4 outputs passed exact geometry, frame count, duration, 24 fps, 32 kHz stereo, and two-channel checks: 10s=240f at 480x640, 15s=360f at 480x640, 20s=480f at 480x640, and 30s=720f at 736x416. No execution error, interruption, NaN/Inf report, Decode failure, Assembly failure, or OOM occurred.
|
||
- Memory acceptance has limited headroom: highest observed device-wide VRAM use was 15.73 GiB and lowest available system RAM was 1.63 GiB. This is a correctness PASS on the tested 16 GiB system, not a guarantee against OOM with other wrappers, model variants, background GPU use, or larger inputs.
|
||
- Evidence is fixed at `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-easy-phase3-gpu-gate-20260830_202000`. `PHASE3_GPU_GATE_REPORT.md` is the human report; `gpu/gate_summary.json`, seven submitted prompts, seven histories, resource samples, media paths, and SHA-256 values are the machine-readable source of truth. The queue was empty after completion.
|
||
- At this gate, `examples/workflows/MiniMax_H3_Continuum_Easy_V38.json` was a 26-node candidate. It has since been superseded by the 27-node Easy Load Audio / Driving Audio candidate recorded above; neither form is final.
|
||
- Phase 3 validation is Easy-focused `45 passed`, latest full source `682 passed`, 195-entry Manifest PASS, runner Python syntax PASS, and `git diff --check` PASS with only the known `pynvml` warning. Accepted source snapshot: `pre-v38-easy-phase3-gpu-gate-approved-excluding-pytest-cache-20260830_201345-0a7c670` (666 files, path/hash mismatch zero, excluding `.git` and the ACL-inaccessible `.pytest_cache`; revision `0a7c6703b7715689c7aab4177e44e5fcd318fe5b`). Production Sampling, package version, runtime code, commit, push, and release remain unchanged.
|
||
|
||
## V3.8 Easy Mode v1.1 / Phase 2 CPU implementation (2026-08-30)
|
||
|
||
- The approved V3.8 Easy Mode contract is fixed as a facade over the unchanged `H3ContinuumSamplerV37` Production path. The public node ID is `H3ContinuumEasyV38` and the display name is `H3 Continuum Easy V3.8`; V3.7 and every older node remain registered and unchanged.
|
||
- Required Easy sockets are MODEL, CLIP, Video VAE, SAMPLER, and SIGMAS. Audio VAE and Scheduler are not Easy inputs. The workflow keeps Core `KSamplerSelect -> SAMPLER` and Core `BasicScheduler -> SIGMAS`; Audio VAE remains on the external Core audio decode path and will be introduced only for a later Reference/Driving Audio helper.
|
||
- Duration is an integer 4-480 seconds with default 10. The resolver uses `chunks=min(16, ceil(duration/10))` and `chunk_seconds=duration/chunks`, so 10/15/20/30 seconds resolve to 1x10, 2x7.5, 2x10, and 3x10. The existing Terminal Merge contract remains 5-second-chunk-only; these approved Easy FL2VA durations deliberately use the existing non-merged Production path.
|
||
- Resolution uses decimal MP (`1 MP = 1,000,000 pixels`), 32-pixel alignment, Draft 0.30 MP, Balanced 0.60 MP, Native 768, and Custom 0.10-16.00 MP. Auto uses First Image aspect or falls back to 16:9. Native preserves aspect, targets short edge 768, and gives the 1344 long-edge cap priority for extreme ratios. Custom is not clamped to Native and reports diagnostic-only Above Native / high-MP warnings.
|
||
- Seed keeps the existing numeric Production contract. The Easy frontend maps only `Randomize` and `Fixed` onto ComfyUI's native `control_after_generate` widget and does not create a backend RNG. Custom MP and Run Name are shown conditionally; Run Storage defaults to Off and retains `Save + Auto Resume` as Advanced.
|
||
- `H3ContinuumEasyReferences` / `H3 Continuum Easy References` is the Phase 2 still-image helper. It has three permanent optional IMAGE sockets (`reference_image_1` through `reference_image_3`) and one `H3_CONTINUUM_EASY_REFERENCES` output. The Easy node expands that private bundle into the same three V3.7 arguments without resizing or conditioning in the helper. Missing slots remain `None`; the existing V3.7 `prepare_reference_assets` path owns gap compaction, Match Output resizing, conditioning, and contiguous `<Picture N>` numbering.
|
||
- Driving Audio, Driving Audio VAE, Reference Audio, Reference Audio VAE, Video Guide Frames, Still Image Guide, and Timeline Video remain outside the Easy Phase 2 schema. Production Sampling and the underlying V3.7 Reference implementation are unchanged.
|
||
- New CPU coverage fixes Resolver values, V3.7 argument mapping, prompt/token passthrough, registration, output contract, Seed frontend behavior, the non-merged 10/15/20/30-second FL2VA policy, no/one/two/three/gapped Reference inputs, First/Last hybrid labels, output geometry, V3.7 mapping parity, and `<Picture N>` behavior. Focused validation is `41 passed`, related Reference/Audio/Timeline validation is `100 passed`, and full source validation is `677 passed`, with only the known `pynvml` warning. All 193 Manifest entries, changed Python syntax, and `git diff --check` pass.
|
||
- Accepted pre-change snapshot: `pre-v38-easy-phase2-approved-excluding-pytest-cache-20260830_170711-0a7c670` (659 files, exact relative-path match, excluding `.git` and the inherited ACL-inaccessible `.pytest_cache`; source revision `0a7c6703b7715689c7aab4177e44e5fcd318fe5b`). The standard snapshot attempt stopped only on `.pytest_cache` before editing.
|
||
- Reference-helper pre-change source snapshot `pre-v38-easy-reference-images-phase2-approved-20260830_184252-0a7c670` contains 666 files with an exact relative-path match, excluding `.git` and `.pytest_cache`. The exact three runtime files changed were snapshotted as `pre-v38-easy-reference-images-runtime-sync-20260830_185057-runtime` before synchronization.
|
||
- ComfyUI_W runtime sync is complete for the changed `nodes.py`, `v3/easy_nodes.py`, and focused test with exact source/runtime SHA-256 equality. After restart, live `object_info` reports `H3 Continuum Easy References`, its three fixed optional IMAGE inputs, `H3_CONTINUUM_EASY_REFERENCES` output, and the new Easy `references` optional input; First/Last remain present and the queue is empty.
|
||
- Runtime pre-change snapshot `pre-v38-easy-runtime-w-sync-approved-20260830_181803-runtime-370` contains 365 files with an exact relative-path match, excluding `.git` and `.pytest_cache`. ComfyUI_W was already a partial runtime package whose 190-entry runtime-local Manifest had 56 pre-existing missing/hash differences versus its own file set. The source 193-entry Manifest was therefore not imposed on the partial runtime; the snapshotted runtime-local Manifest was restored after diagnosis. Source Manifest verification remains 193/193 PASS.
|
||
- Phase 2 did not add the Minimal Workflow or any Audio/Video Reference helper, run a GPU gate, update package version, commit, push, or release.
|
||
|
||
## E2-8 Second Pass contract (2026-08-22)
|
||
|
||
- V3.4 remains the stable generation contract; UI, sockets, sampling, Run Storage identity, and Assembly behavior are unchanged.
|
||
- `assembly_plan.second_pass_contract` is additive V3.5 metadata organized by physical decode group.
|
||
- Prompt source is the stored physical prompt; a future Second Pass node must not reparse Sequence Prompt.
|
||
- Video B/C/T and physical group count/order are fixed; only latent H/W may be preserved or enlarged.
|
||
- Audio output is the original physical audio LATENT passed through bit-exact; refined audio is not adopted in V1.
|
||
- E2-8A Contract, E2-8B1 CPU/static validation, E2-8B2 external LBH GPU PoC, and E2-8C Second Pass GPU acceptance are formally PASS. The E2-8C CPU baseline is `260 passed`.
|
||
- External LATENT nodes may discard custom dictionary metadata. Identically shaped group reordering therefore cannot be cryptographically detected; count, position, T, geometry, and audio-shape checks are enforced.
|
||
|
||
Updated: 2026-08-28
|
||
|
||
## V3.7.0 Conditioning Adapter and RefineSchedule release (2026-08-30)
|
||
|
||
- Post-review frontend/input hardening is complete. `H3ContinuumSamplerV37` now follows the V3.6 compact frontend path for project identity, hidden internal widgets, Run Storage conditional controls, queue normalization, runtime diagnostics/debug/preview settings, and permanent Reference Audio label normalization. Still Image Guide now rejects every IMAGE batch except exactly one image at both public payload creation and defensive source preparation.
|
||
- Guide/First/Last endpoint collisions are regression-covered at physical frame `0` and `123`: the explicit Still Guide role marker takes priority over frame-number inference, so the owned Guide source is reencoded rather than the First/Last image. Reference Audio remains the accepted V3.5.1 permanent-socket design; the withdrawn frontend `Hidden / Show` widget was not restored.
|
||
- V3.7 has an internal, versioned `RefineSchedule` contract with `External`, `Full`, `Tail`, and `Partial` modes. `External` preserves the exact incoming SIGMAS tensor; Tail/Partial use exact source slices. The contract records source/effective sigma hashes, evaluation range, video-random/audio-locked noise policy, and is included in `second_pass_contract` identity without invalidating reusable First Pass Run Storage.
|
||
- Hi-Res Fix and Second Pass expose the schedule only as an optional Python-internal argument. Their public `INPUT_TYPES`, saved-workflow schema, V3.6/V3.5 nodes, Production defaults, and existing SIGMAS socket remain unchanged. Tail 10 and Tail 6 are reproducible from the same 10-evaluation source schedule.
|
||
- Still Image Guide now owns a CPU clone and SHA-256 identity of the original Guide source image. Only its owner physical group carries that source, and Second Pass rebuilds the marked Guide keyframe from the original image at target latent geometry; First/Last, Terminal Merge ownership, Core PackedLayout/RoPE construction, and First Pass audio bit-exact passthrough remain unchanged.
|
||
- GPU Correctness PASS: internal Tail 6 at 704x704 used the exact six-evaluation suffix. Under the same Sage path as the earlier external `SplitSigmas` Tail 6 baseline, decoded Video SSIM is `0.982936` and decoded PCM SHA-256 is identical; both are 120 frames / 5.000 seconds. The 1152x1152 Guide run encoded the First Image plus the original 1254x1254 Guide source to 72x72 latent geometry and completed without node/finite/geometry/AV errors. It remains low-motion/freeze-detected, consistent with the already-known hard-anchor quality limitation; Still Image Guide Production promotion therefore remains **HOLD**.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v37-pillars-gpu-gate-20260830`. Accepted snapshots: `pre-v37-refine-schedule-conditioning-approved-ecc72da-accepted-20260830_005156` (source, 641 files, zero SHA mismatch) and `pre-v37-refine-schedule-conditioning-runtime-w-approved-accepted-20260830_005202` (ComfyUI_W runtime, 353 files, zero mismatch).
|
||
- Final Release Gate is `633 passed`; the 185-entry Manifest, Python/JavaScript syntax, `git diff --check`, ComfyUI_W runtime verifier, live V3.7/V3.6 registration, empty queue, and source/runtime SHA-256 equality for 25 release/Production files pass. ComfyUI_W restarted on 0.34.2 and reports `H3 Continuum 3.7.0 loaded`. Runtime logic and Production defaults are unchanged. Release commit `fe4ff9c20c2cc8bb375625d1534f5673a737d1be`, annotated tag `v3.7.0`, and the GitHub Release were published on 2026-08-30.
|
||
- A post-release English README clarification is staged locally: Tail 10 `100.27 s` versus Tail 6 `64.18 s` (`36.0%` lower), the separate Learned Latent measurements, and the distinction between RefineSchedule contract management and the speed gained by six rather than ten evaluations. The V3.6 hero was replaced by a V3.7 Tail 6 infographic, and a V3.7 Sampler UI image with the optional `guide` input was added. The Manifest now has 187 verified entries; runtime logic, version, tag, and Production defaults are unchanged. Commit and push are pending.
|
||
- The original V3.6 Masked AV infographic was preserved as `docs/images/v36-masked-av-continuation.png` and restored directly below the `V3.6 Masked AV Continuation` heading. The V3.7 hero and V3.7 Sampler UI remain in their V3.7/current-interface positions. The 188-entry Manifest and full `633 passed` regression pass; runtime logic, version, and release tag are unchanged. Commit and push are pending.
|
||
- `examples/workflows/MiniMax_H3_Continuum_V37.json` and its ZIP are now the recommended V3.7 starting workflow. They are exact V3.6 graph derivatives with only Sampler node `305` and its S&R class changed from `H3ContinuumSamplerV36` to `H3ContinuumSamplerV37`; all links, widgets, Standard backend, Hi-Res Fix OFF default, Assemble settings, and Production defaults are unchanged. The optional Still Image Guide is deliberately disconnected because it remains Experimental / Production HOLD. V3.6/V3.5 workflows and the untracked PDD workflow are preserved. Static workflow/ZIP checks and the full `636 passed` regression pass; the Manifest has 190 verified entries. No GPU rerun, version change, commit, or push was performed.
|
||
|
||
## Sampling layout-validation measurement (2026-08-28)
|
||
|
||
- Phase 1 Correctness PASS snapshots `phase1-correctness-pass-source-accepted-20260828_172802-ecc72da` and `phase1-correctness-pass-runtime-w-accepted-20260828_172802-ecc72da` are the formal optimization A/B baseline.
|
||
- Detailed Report now measures the existing `validate_native_continuity_layout()` call count, full-validation count, status counts, cumulative host wall including its existing device-to-host transfers/implicit synchronization, physical-group step contribution, and observable accelerator markers. It does not add `torch.cuda.synchronize()` and is inactive outside Detailed Report. Validation, Sampling, payload, schema, workflow, metadata, and public UI contracts are unchanged.
|
||
- The live ComfyUI_W / RTX 5060 Ti matrix used I2VA + Continuation, Compatibility Reference Context, 2x5 seconds, 384x384, 4 steps, fixed seed `361040`, and identical prompts. Sage and Sage+Sol each made `4` full calls (`1.000`/validated step) at `0.472` and `0.530 ms/step`; Sage+Spectrum and Sage+Sol+Spectrum each made `8` full calls (`2.000`/validated step) at `1.164` and `1.108 ms/step`. Every route reported zero validation failures. Global `--use-sage-attention` is not exposed through the inspected MODEL markers, so Sage route identity is taken from the controlled workflow graph and recorded backend launch arguments rather than inferred from `accelerator_summary()`.
|
||
- The representative Sage run used 576x576, 2x5 seconds, 20 steps, and the same seed. It made `20` full calls, `1.000` call/validated step, with `14.289 ms` cumulative validator wall (`0.714 ms/step`) versus `321.632848 s` measured Sampling and `382.055317 s` API elapsed. The validator therefore accounts for approximately `0.00444%` of Sampling and `0.00374%` of API time.
|
||
- Decision: do not implement a layout-validation cache or fast-path. Even eliminating the measured validator completely has a maximum representative saving of about `14 ms`; it is not a material Sampling optimization target. Spectrum doubles the validator call count in this matrix but remains about `1.1 ms/step`. No Sampling optimization was started. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\layout-validation-profile-phase`.
|
||
|
||
## V3.6.1 Run Storage correctness hotfix (unreleased, 2026-08-28)
|
||
|
||
- Absent Last Frame hashes are canonicalized as `""` for every new chunk contract. Candidate revisions may bypass a hash mismatch only when both top-level Last Frame values are absent and normalizing the per-chunk `last_frame_hash` is the sole contract difference.
|
||
- This narrow path reuses legacy V3.6.1 final chunks stored with `"none"`; any real Last Frame or any other field difference remains incompatible. Run Storage schema and Sampling Contract versions are unchanged. FL2VA Long Terminal Merge atomicity remains unchanged.
|
||
- Compact UI now resets `Regenerate From = Auto` and `Variation Nonce = 0` whenever Run Storage is Off, including graph reload and final API serialization. Backend validation remains enabled.
|
||
- CPU/static regression expects normal no-Last-Frame I2VA 2→3 at 2/1, legacy `none` at 2/1, real Last Frame at 1/2, Terminal Merge at 1/2, exact resume at full reuse, and frontend Off/reload/queue normalization. Focused validation is `64 passed`; full validation is `565 passed` with only the known `pynvml` warning.
|
||
- Actual ComfyUI_W 0.34.2 / RTX 5060 Ti 16 GiB GPU regression copied only legacy revision `9c35c64b529a5c2c` into isolated test Run `run_95c2caec1a0846a09cf3570acb29e27f`. Prompt `bf619577-2057-4635-8fef-2537a40295c4` completed as `2 reused, 1 generated`, first regenerated Chunk 3, output `MiniMax_H3_00008_.mp4`. Exact 2-chunk resume prompt `e593b06a-9a37-4cc1-932c-0d13e55979c1` completed as `2 reused, 0 generated` with no regenerated chunk.
|
||
- G3 is repaired without changing Run Storage identity or Sampling behavior: the `FLF_STRATEGY` compatibility guard now applies only when Terminal Merge is actually enabled. Non-Terminal FL2VA therefore retains a prefix already validated by Run Storage; Terminal Merge still rejects a missing/mismatched strategy before its atomic pair handling. Focused validation is `63 passed`; full validation is `568 passed` with only the known `pynvml` warning.
|
||
- **Phase 1 GPU Acceptance G1-G6 is PASS.** G1 canonical no-Last-Frame I2VA 2→3 and G2 an isolated copy of actual legacy V3.6.1 `"none"` cache both completed at `2 reused / 1 generated`, Sampling 1. G3 real-Last-Frame non-Terminal FL2VA completed at `1 / 2`, Sampling 2, and no longer reports `FLF chunk contract changed`. G4 Terminal Merge remains `1 / 2` with one physical `77T Video / 433T Audio` Sampling call. G6 Run Storage Off completed with one Sampling call and no queue error.
|
||
- G5 fresh 3-chunk and matched 2→3 extension remain bit-exact: all three raw Video latent SHA-256 values, all three raw Audio latent SHA-256 values, decoded RGB24 SHA, decoded PCM SHA, 360 frames, and 480256 PCM samples/channel are identical. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\phase1-gpu-acceptance-20260828\20260828_170034\gate_summary.json`.
|
||
- Verified pre-G3 snapshots are source `pre-g3-fl2va-prefix-correctness-source-accepted-20260828_165400-ecc72da` (632 files excluding `.git` and the inaccessible `.pytest_cache`) and ComfyUI_W runtime `pre-g3-fl2va-prefix-correctness-runtime-w-accepted-20260828_165400-ecc72da` (351 files excluding the inaccessible `.pytest_cache`), both with zero SHA-256 mismatch. Final accepted snapshots are `phase1-correctness-pass-source-accepted-20260828_172802-ecc72da` and `phase1-correctness-pass-runtime-w-accepted-20260828_172802-ecc72da`. All 180 Manifest entries, source/runtime SHA-256, runtime verifier, Python/JavaScript syntax, empty queue, and `git diff --check` pass. Public UI, metadata, workflows, Reference/Masked transport identity, schema/version, and optimizations are unchanged; Sampling hot-path optimization remains a separate later phase.
|
||
|
||
## V3.6.1 Continuity fallback hotfix (2026-08-27)
|
||
|
||
- V3.6 Standard now resolves its continuation transport before Production execution and Run Storage identity creation. With Audio Continuity ON, Balanced 22 remains `masked_av_prefix_22_v1`; Fast 5, Strong 39, and Auto resolve to the accepted legacy `reference_context_v1` transport instead of reaching the incompatible 22-frame Masked AV contract.
|
||
- Standard with Audio Continuity OFF remains the dynamic `masked_video_prefix_v1` path for 5/22/39-frame Continuity. Compatibility remains `reference_context_v1` for every mode. User Backend and Continuity widget values are never rewritten.
|
||
- Non-Balanced Standard fallback reports `Continuation Backend: Standard`, `Resolved transport: Reference Context`, and the reason before the normal status. Run Storage follows the resolved transport: fallback and explicit Compatibility safely share the identical legacy Reference identity, while Balanced Masked AV stays revision-separated.
|
||
- No V3.5.3/V3.4 routing, Sampling, Conditioning payload, Terminal Merge, Assembly, Seam, Run Storage schema, Node ID, socket, or workflow serialization contract changed. Dynamic 5f/39f Masked AV remains a separate V3.6.2-or-later research item.
|
||
- Focused routing/Masked/Run Storage/Prompt/release validation is `108 passed`; final full source pytest is `543 passed`; all 178 Manifest entries, compileall, and `git diff --check` pass with only the known `pynvml` FutureWarning and inaccessible `.pytest_cache` listing notice. `v2/prompts.py`, `v3/driving_nodes.py`, and `version.py` are synchronized to ComfyUI_W with exact SHA-256 equality, and the V3.6.1 runtime verifier passes.
|
||
- The representative GPU Smoke is PASS on ComfyUI_W 0.34.0 / RTX 5060 Ti 16 GiB with Sage Attention and `--disable-pinned-memory`: public `H3ContinuumSamplerV36`, Standard, Audio Continuity ON, Strong 39, T2VA 2x5 seconds, 576x576, 20 steps. Prompt `f5d5c36c-fa5b-4031-ba0e-349cf70ff64f` completed in `412.034s`; the report begins with the expected `Resolved transport: Reference Context` and fallback reason, retained `124f + (158f - 39f) = 243f`, and reported no execution, finite-validation, Decode, Assembly, or save failure. Output `D:\output\video\comfy_video\video\V361_Strong39_Fallback_Smoke_00001_.mp4` is `576x576`, `240f`, `24fps`, exactly `10.000s`, with 32 kHz stereo audio of the same duration.
|
||
- Accepted pre-change snapshot: `pre-v361-continuity-fallback-source-accepted-20260827_194154-e626053` (598 files excluding `.git` and the known ACL-inaccessible `.pytest_cache`, source revision `e62605389302192394446645cf3483fff08dfbf2`).
|
||
- Accepted runtime snapshot: `pre-v361-runtime-w-sync-accepted-20260827_195452` (350 files excluding `.git` and the known ACL-inaccessible `.pytest_cache`).
|
||
|
||
## V3.6.1 Prompt Parser hotfix implementation (unreleased, 2026-08-27)
|
||
|
||
- Shared `v2/prompts.py` now detects mixed Timeline headers and standalone `---` List separators without stopping generation. Auto and explicit Timeline keep Timeline assignment, remove only standalone separator lines, emit `H3C-P105`, and retain the existing `H3C-P101` previous-prompt fallback for uncovered chunks. Text after a removed separator is not reassigned to another chunk.
|
||
- Explicit List retains its existing separator parsing and only reports `H3C-P105` when valid Timeline headers are present. Explicit Fixed remains literal and warning-free. Invalid Timeline still fails open to the original Fixed text with `H3C-P100`.
|
||
- Reports with parser warnings now begin with `PROMPT PREFLIGHT WARNING`; Timeline source/fallback diagnostics and explicit-List section mappings expose the resolved per-chunk assignment. Prompt Plan Preview uses the same Python parser and report. No Frontend/JavaScript parser or Toast was added.
|
||
- Run Storage schema is unchanged. A dedicated regression confirms that hashes from the legacy mixed Prompt text and the normalized resolved Prompt produce distinct revision identities, preventing incorrect reuse.
|
||
- Validation: parser plus Run Storage focused `71 passed`; related public/schema/workflow suite `94 passed`; full pytest `538 passed`; compileall, V3.6.0 source runtime verifier against ComfyUI_WAN, and `git diff --check` pass. Only the known `pynvml` FutureWarning appeared. GPU testing was intentionally not run.
|
||
- Accepted pre-change snapshot: `pre-v361-prompt-parser-hotfix-accepted-20260827_190257` (214 Git-visible/untracked nonignored files, SHA-256 mismatch 0, revision `0d18e3abe0923502171f47f24769e957057eaf15`). Frontend remains unchanged. This parser hotfix is now included in the local V3.6.1 metadata and synchronized ComfyUI_W runtime together with the Continuity fallback; commit and push remain pending.
|
||
|
||
## V3.6.0 public release implementation (2026-08-27)
|
||
|
||
- Added the new public `H3ContinuumSamplerV36` / `H3 Continuum Sampler V3.6` without replacing or redirecting V3.5.3/V3.4 nodes. Its final required widget is `Continuation Backend` with only `Standard` and `Compatibility`; internal transport identifiers are not exposed.
|
||
- `Standard` selects accepted Masked AV when Audio Continuity is ON and accepted Masked Video when it is OFF. `Compatibility` selects the original Reference Context route. Existing V3.5.3 remains on Reference Context with its prior schema and widget sequence.
|
||
- Added `examples/workflows/MiniMax_H3_Continuum_V36.json` as the recommended template, retaining the V3.5 template unchanged. Version/package metadata and English/Japanese release documentation are updated to 3.6.0; V3.5 Second Pass, Hi-Res Fix, Conditioning Bridge, V3.5 Assembly, and older workflows remain available.
|
||
- The public-node GPU Smoke used `H3ContinuumSamplerV36` directly through the normal API, not the private diagnostic node: Standard backend, 2x5-second I2VA + Reference Image + Reference Audio, 384x384, four steps, Audio Continuity ON, Audio Seam Auto. Prompt `1cff2ce7-6587-4d71-903d-b04bb304fdd5` completed successfully in `65.21s`; the report confirms a 22f/37T masked AV target prefix on chunk 2.
|
||
- Smoke output is `D:\output\video\comfy_video\video\V36_Release_Smoke_Reference_AV_Standard_00001_.mp4`: finite 384x384, 240 frames, 24 fps, 10.000-second video, and 32 kHz stereo 10.000-second audio. Evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v360-release-smoke\20260827_172835`.
|
||
- Source focused release validation is `81 passed`; final full source validation is `527 passed` with only the known `pynvml` FutureWarning. V3.6.0 WAN runtime verifier, JavaScript syntax, changed-file `py_compile`, the 178-entry Manifest, README link validation, and source/runtime SHA-256 equality for 12 Production files all pass. The live API queue is empty and the V3.6 backend remains running.
|
||
- Accepted snapshots are `pre-v360-public-release-accepted-20260827_171442` (212 Git-visible source files) and `pre-v360-public-runtime-wan-20260827_172531` (the five runtime files changed for public registration/version/UI). Final repository validation is complete; commit/push/release are the remaining publication actions.
|
||
|
||
## V3.6 Production Integration Gate (2026-08-27)
|
||
|
||
- PIG-0 through PIG-5 are complete. The Masked AV backend and V3.6 public/default promotion are **PASS**. During the gate, published/default V3.5.3 remained `reference_context_v1` and no public schema changed; the separately authorized V3.6 UI/version implementation is recorded above. V3.5.3 and V3.4 behavior remains unchanged.
|
||
- Run Storage now scopes only nondefault continuation transports. `reference_context_v1` retains its exact legacy identity/revision behavior, while `masked_av_prefix_22_v1` records `continuation_transport` in both the identity and Sampling Contract. Reference and Masked chunks therefore cannot be mixed.
|
||
- The integration gate also found and fixed a pre-existing resume defect: temporary Run Storage Sessions did not restore Terminal Merge `execution_semantics`, so an apparently valid prefix was discarded before Sampling. Resume Sessions now carry the stored execution semantics, and the reported reuse count is synchronized after atomic terminal-pair pruning.
|
||
- Real-cache PIG-2 PASS evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\v36_pig2_masked_20260827_164303`. Initial Save used `0 reused / 3 generated`; Auto Resume used `3 / 0` with zero Sampling calls; Regenerate From Chunk 2 and Chunk 3 each used `1 / 2` and exactly one `77T Video / 433T Audio` terminal physical Sampling call. Neither request reused or sampled only one logical half of the terminal pair.
|
||
- PIG-1 CPU coverage proves that Reference Image and standalone Reference Audio remain present exactly once in every normal and Terminal physical group under Masked AV, with no old continuation Reference block. PIG-3 then completed a representative 2x5-second I2VA + Reference Image + Reference Audio GPU run at 384x384 / 4 steps. Both physical groups contained image/audio references, protected Video/Audio prefixes finalized bit-exact, and the final output is finite 240 frames / 24 fps / 10.000 seconds with 10.000-second stereo audio.
|
||
- PIG-3 evidence is `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-production-integration\pig3-reference-masked`; output video is `D:\output\video\comfy_video\video\V36_PIG3_Reference_Image_Audio_Masked_AV_00001_.mp4`. At the retained 124-frame boundary with existing Audio Seam Auto, the sample jump is `0.004261`, below the local median (`0.005129`) and only `0.0876x` the local p99; PCM is finite. This representative numeric gate is PASS, but prior R2 seeds show mixed Seam-Auto improvement, so no universal click-removal claim is made without listening/generalization.
|
||
- PIG-5 policy is fixed: V3.6 uses Masked AV as the standard continuation backend and retains Reference Context only as an Advanced compatibility fallback. Raw transport implementation names are not exposed in the UI. Run Storage identity remains transport-scoped, Terminal Merge regeneration remains atomic per physical group, and Audio Seam Auto remains an independently evaluated post-Sampling feature. This gate preceded the separately recorded V3.6 public UI/default implementation above.
|
||
- The accepted PIG-3 output passed subjective listening: no click, dropout, unnatural stereo positioning, or other audible boundary defect was reported. PIG-5 and V3.6 public-promotion eligibility are therefore **PASS** without reopening the completed structural, cache, Reference, Terminal Merge, or Audio Seam numeric gates.
|
||
- Final validation is `61 passed` focused and `521 passed` full pytest with only the known `pynvml` FutureWarning. Changed-file `py_compile`, V3.5.3 WAN runtime verification, source/runtime SHA-256 equality for `run_storage.py`, `v2/sequence.py`, and the diagnostic node, `git diff --check`, and empty API queue all pass.
|
||
- Accepted pre-change snapshots are `pre-v36-production-integration-pig0-accepted-20260827_161517` and `pre-v36-production-integration-pig2-diagnostic-accepted-20260827_162943`. Commit/push were not performed, and the ComfyUI_WAN backend remains running.
|
||
|
||
## V3.6-R2 FL2VA Long Terminal Merge multi-seed GPU gate (2026-08-27)
|
||
|
||
- The shortened final gate is **PASS for the private GPU research path**. Published/default V3.5.3 remains `reference_context_v1`; no public node/schema, saved-workflow contract, V3.4 behavior, Sampling implementation, Assembly, Seam, or Run Storage behavior changed.
|
||
- Accepted evidence consists of Seed `238985343135606` full Gate A, Seed `238985343135607` full `reference_context_v1` versus `masked_av_prefix_22_v1` comparison plus cache-only Audio Seam OFF derivatives, and Seed `238985343135608` Masked AV structure/practical acceptance. The redundant Seed 608 Reference run was intentionally omitted.
|
||
- Every measured run retained 3 logical chunks and 2 physical groups. Group 1 is `124f / Video 37T / Audio 207T`; Terminal Group 2 is logical `[2,3]`, `paired_timeline_v1`, `terminal_merged=true`, trim `22f`, `260f / Video 77T / Audio 433T`. Physical seeds were `15439951268068580545`, `4181078118376080811`, and `6961917749643595863` for Seeds 606–608. Final outputs are finite `640x640`, `360f`, `24fps`, `15.000s`, with 15.000-second audio.
|
||
- Seed 606 and 607 Group 1 Video/Audio hashes are identical between Reference and Masked routes. For all three Masked runs, Core raw protected-prefix drift was at most `2.384185791015625e-07` Video and `5.960464477539063e-08` Audio; scoped final restore produced bit-exact 7T Video and 37T Audio prefixes with SHA-256 equality and maximum difference `0`. Generated regions were not restored.
|
||
- Terminal Group 2 PackedLayout was `35,865` rows for Reference and `32,991` for Masked, removing exactly `2,874` Reference rows (`8.01%`). CUDA-synchronized Group 2 Sampling was `574.134 -> 529.933s` for Seed 606 and `560.782 -> 544.413s` for Seed 607. Both paired comparisons favor Masked; paired percentage median is `5.31%`. Reference/Masked time medians are `567.458s` and `544.413s`, respectively (`4.06%` lower by median comparison). Seed 608 Masked was `549.704s`.
|
||
- Across the five main runs, monitored process peak RSS median was `9.674 GiB`. Maximum observed whole-device usage was `15.646 GiB`; this is an order-sensitive device-used monitor, not only Continuum allocation. Terminal-group CUDA allocator peak was about `4.06 GiB` for Reference versus `3.80 GiB` for Masked. No OOM, NaN/Inf, or retained-GPU-prefix regression occurred.
|
||
- Contact sheets covering nominal 5/10/15-second positions and actual retained boundaries `124f/243f` show no obvious flash, pose collapse, direction reversal, or temporal break in Seeds 606–608. Decoded audio is finite, exactly 15 seconds, and has no >=200 ms dropout below -50 dB. Existing Audio Seam Auto strongly improved Seed 606 Masked jumps (`0.035014 -> 0.002018` at 124f; `0.004495 -> 0.003130` at 243f), but Seed 607 was mixed (`0.059640 -> 0.133805` at 124f; `0.021297 -> 0.019704` at 243f). Therefore no universal click-removal claim is made from numerical jump alone.
|
||
- Final validation is `516 passed` with only the known `pynvml` FutureWarning, after rerunning with a fresh writable basetemp because the inherited global pytest temp root denied access. V3.5.3 WAN runtime verification, source/runtime SHA-256 for seven synchronized Production/diagnostic files, `git diff --check`, and empty API queue pass. The backend remains running as requested.
|
||
- **Production promotion: HOLD.** FL2VA/Terminal Merge structure, multi-seed execution, integrity, and performance are accepted, but Reference conditioning variants, Run Storage/resume identity, public UI/default-policy decisions, and broader audible Audio Seam acceptance remain outside this gate. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-fl2va-combined-20260826\gate-a` and `...\r2-short`. Pre-change snapshot: `pre-v36-r2-short-gate-20260827_000840`. Commit/push were not performed.
|
||
|
||
## V3.6 shared chunk-duration range (2026-08-27)
|
||
|
||
- V3 Production, V3.4/V3.5 inherited schemas, the V2 compatibility sampler/facade, and Prompt Plan Preview now share `chunk_seconds`: default `5.0`, minimum `4.0`, maximum `30.0`, step `0.1`.
|
||
- One `chunk_seconds` value still applies to every logical chunk. Per-chunk variable durations were not added; Assembly, Timeline, Continuation, and Run Storage retain the shared-duration contract and existing workflow inputs remain compatible.
|
||
- `5–15s` remains the recommended and validated range. `15–30s` is usable rather than hidden/experimental, with a shared tooltip warning that high resolution can substantially increase VRAM use and processing time.
|
||
- Prompt Plan creation and validation accept `30.0` and reject values above it. Timeline mapping was verified for `[0-30s]` / `[30-60s]`. Temporal extension tests now cover the complete `4.0–30.0s` UI range and preserve the existing native H3 `17k+5` grid.
|
||
- The only remaining `max: 15.0` is the separate legacy V1 `H3ContinuumJoin.extend_seconds` control; it is not `chunk_seconds`, is not used by the V2/V3 Production path, and was intentionally left unchanged.
|
||
- CPU/static acceptance is PASS: focused duration and inherited-schema validation `89 passed`; canonical full pytest `516 passed`; changed-file `py_compile`, `git diff --check`, runtime verifier, source/WAN SHA-256, live API schema, and empty queue all pass. ComfyUI_WAN was restarted with Sage Attention and `--disable-pinned-memory` and remains running.
|
||
- A validation-only GPU launch used one T2VA chunk, `chunk_seconds=30.0`, `384x384`, one Sampling step, and fixed seed `238985343135609`. API prompt `9e665f85-b32f-4b17-9734-219fb284e087` was accepted with no node errors and entered H3 model initialization/Sampling without duration, shape, Prompt Plan, immediate OOM, or NaN/Inf errors. Per user scope, it was interrupted after `31.87s` rather than decoded or saved; this is an execution-entry smoke PASS, not a completed 30-second quality/performance acceptance. Queue returned to `0/0`.
|
||
- The previously prepared FL2VA joint multi-seed run was cancelled at the user's request. Only its one-step cache warm-up completed; the first measured run was interrupted before node execution and is not acceptance evidence.
|
||
- Accepted snapshots are `pre-v36-chunk-duration-30-source-20260826_234716` (209 Git-visible source files, zero missing/mismatch) and `pre-v36-chunk-duration-30-wan-runtime-20260826_235423` (four exact runtime files). Commit/push were not performed.
|
||
|
||
## V3.6-R2-2 Balanced 22f / 37T Joint AV GPU gate (2026-08-26)
|
||
|
||
- R2-2 is **PASS as a one-seed GPU structure/performance gate with the existing Audio Seam Auto path**, not a Production promotion. The accepted pair used T2VA 2x5 seconds, 640x640, Balanced 22, Audio Continuity ON, fixed seed `238985343135605`, Prompt/CLIP cache HIT, CUDA synchronization, Sage Attention, `--disable-pinned-memory`, Exact Duration ON, and no First/Last/Reference/Video Guide/Run Storage/Spectrum/Sol/Video Seam.
|
||
- Added only the private `masked_av_prefix_22_v1` research transport. It preserves the existing Balanced contract: Video prefix `22f = 7T = 0.916667s`, Audio prefix `37T = 0.925s`, an intentional mask-boundary asymmetry of `1/120s = 8.333ms`, target `141f / Video 42T / Audio 235T`, trim `22f`, and net new `119f`. Public V3.5.3/V3.4 schemas and defaults remain unchanged.
|
||
- Masked finalization is exact: source Video tail and finalized target Video prefix are bit-exact with max diff `0`; the same holds for the Audio 37T tail/prefix. Core raw output before the scoped restore differed by at most `2.384185791015625e-07` for Video and `5.960464477539063e-08` for Audio. Only the protected prefix regions are restored; generated regions remain sampler-owned.
|
||
- Reference continuation used `20,205` packed rows (`2,800` video-reference rows plus `74` audio-reference rows); Masked AV used `17,331` rows with no reference block or old Spectrum interop hint. The `2,874`-row reduction is `14.22%`. Synchronized continuation Sampling was `226.373s -> 206.437s`, an `8.81%` reduction; whole API time was `435.907s -> 420.752s`, a `3.48%` reduction. This is one-seed evidence, not a universal performance claim.
|
||
- Both routes produced finite `640x640`, `240f`, `24fps`, `10.000s` outputs. Eight-frame boundary contact sheets show no obvious pose snap, direction reversal, camera collapse, NaN/Inf, or masked-prefix visual regression.
|
||
- With Audio Seam OFF, Reference boundary jump was `0.082725`; Masked AV was worse at `0.268163` (`7.57x` its local p99 and the `99.9989` file-wide percentile). Reassembling the same cached Masked latents with existing Audio Seam Auto took `7.10s` without Sampling and reduced the jump to `0.029944` (about `88.83%` lower, `0.86x` local p99, `94.31` percentile). RMS ratio also moved `1.12196 -> 0.97724`. R2 adds no new Audio Seam algorithm; raw Seam-OFF 22f/37T is not accepted as click-free.
|
||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate2-20260826`; accepted MP4s and the cached Audio-Seam-Auto derivative are under `D:\output\video\comfy_video\video`. Focused tests are `24 passed`; final full pytest is `510 passed`; changed-file `py_compile`, `git diff --check`, source/runtime SHA-256, runtime verifier, and empty queue all pass. Production remains **HOLD** pending multi-seed/audio listening generalization, Reference/FL2VA/Terminal Merge/Run Storage integration, public UI decisions, and final Production acceptance.
|
||
|
||
## V3.6-R2-0 Joint AV masked-prefix CPU/static gate (2026-08-26)
|
||
|
||
- R2-0 is implemented as a private research transport, `masked_av_prefix_39_v1`; the public/default V3.5.3 transport remains `reference_context_v1`. No public widget, Node ID, saved-workflow schema, `v2/sampling.py`, seed, Plan, trim, Decode, Assembly, Seam, Run Storage, or V3.4 contract changed.
|
||
- The exact-duration research case uses Strong 39-frame Continuity. A 39-frame video tail maps to 12 video latent slots, and the same 1.625-second duration maps to 65 audio latent ticks. The next physical target is 158 frames / video `47T` / audio `263T`, trims 39 frames, and retains 119 new frames.
|
||
- The target copies the prior finalized video tail into video `[0:12]` and audio tail into audio `[0:65]`. Full masks preserve those prefixes with zero and generate only video `[12:47]` / audio `[65:263]` with one. Old Context Reference conditioning and the Spectrum reference-context hint are not emitted. The original First Image remains attached only to chunk 1.
|
||
- After Sampling, Continuum restores only the protected video and audio prefixes from the normalized pre-sampling target. CPU regressions require bit-exact video/audio prefix equality and prove that generated video/audio regions are not restored or overwritten.
|
||
- The 39f and 65T durations are equal, but the native H3 audio grid can retain a signed source phase offset. The validated 124-frame source is `+1/3` audio tick (about `+8.33 ms`) and a 158-frame target is `-1/3` tick; this is recorded explicitly rather than described as zero-phase alignment.
|
||
- CPU/static R2-0 is **PASS**: focused masked/temporal validation `21 passed`, canonical full pytest `507 passed`, and changed-file `py_compile` pass with only the known `pynvml` FutureWarning. GPU R2-1 (`Reference Context 39f` versus `Masked AV Prefix 39f`), audio-boundary measurements, runtime synchronization, and Production promotion remain pending.
|
||
- Accepted pre-change snapshots are `pre-v36-r2-0-joint-av-source-accepted-20260826_203022` (571 files, SHA-256 mismatch 0, source revision `67a0d80117bd8332d2768d093dff343001c87d60`) and `pre-v36-r2-0-joint-av-wan-runtime-accepted-20260826_203022` (three exact runtime files). Commit/push were not performed.
|
||
|
||
## V3.6-R2-1 39f / 65T Joint AV GPU structure gate (2026-08-26)
|
||
|
||
- R2-1 is **PASS as a GPU structure/interoperability gate**. It is not a Production promotion. The accepted pair used T2VA 2x5 seconds, 640x640, Strong 39, Audio Continuity ON, the same fixed prompt/seed `238985343135604`, Prompt/CLIP cache HIT, CUDA synchronization around each Sampling call, `--disable-pinned-memory`, Sage Attention, and no First/Last/Reference/Video Guide/Run Storage/Spectrum/Sol/Video Seam. Reference ran before Masked after one excluded warm-up.
|
||
- `reference_context_v1` emitted Core's combined `video_audio` Reference block. Its continuation PackedLayout was `24,317` rows: target text `61`, audio `526`, video `18,800`, plus `4,800 ref_img` and `130 ref_audio` rows. `masked_av_prefix_39_v1` emitted no old refs/interop and used `19,387` rows: the identical target rows only. The measured reduction is exactly `4,930` rows / `20.27%` relative to Reference.
|
||
- CUDA-synchronized Chunk 2 host/Sampling time was `332.460s` for Reference and `241.888s` for Masked (`27.24%` lower). API wall time was `556.058s` versus `465.925s` (`16.21%` lower). This one-seed Gate is an explanatory performance signal, not a generalized speed claim.
|
||
- Masked target construction and masks passed on GPU: Video `[0:12]=0`, generated Video `[12:47]=1`, Audio `[0:65]=0`, generated Audio `[65:263]=1`; source tails matched both target prefixes. Core raw integration differed by at most `2.384185791015625e-07` for Video and `5.960464477539063e-08` for Audio. The finalized physical-group prefixes are bit-exact by SHA-256 and maximum difference `0` for both streams; generated regions are not restored.
|
||
- Both outputs are finite `640x640`, 240 frames, 24 fps, and 10.000 seconds. Boundary contact sheets at frames `112,118,123,124,125,130,136,148,160` show no gross snap, collapse, persistent artifact, or identity/background discontinuity in either route.
|
||
- With Audio Seam OFF, the exact decoded boundary sample is the global maximum single-sample jump for both routes: Reference `0.352289`, Masked `0.287521`. Masked is better but the raw hard boundary is not accepted as intrinsically click-free. The same cached Masked AV latents reassembled with the existing Audio Seam Auto in `7.23s` without re-running Sampling; boundary jump fell to `0.005774` (about `98.0%` lower and only the `37.4` percentile of file-wide sample differences), while DC jump fell `0.002789 -> 0.001103`. Therefore the existing Production seam path resolves the observed hard cut; R2 adds no new Audio Seam algorithm.
|
||
- Chunk 2 allocator peak was lower for Masked (`2,407,722,872` allocated / `2,785,017,856` reserved bytes) than Reference (`2,950,494,614` / `3,590,324,224`). Whole-process/device monitor peaks are order-sensitive and are not accepted as a memory claim from this one pair.
|
||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r2-gate1-20260826`; accepted MP4s are `V36_R2_Gate1_reference_39f_seed238985343135604_00001_.mp4`, `V36_R2_Gate1_masked_av_39f_65t_seed238985343135604_00001_.mp4`, and the cached Audio-Seam-Auto derivative under `D:\output\video\comfy_video\video`. The backend remains running. R2-2 22f / 37T, multi-seed/audio-quality generalization, References, Terminal Merge, Run Storage, public UI, and Production promotion remain pending.
|
||
|
||
## V3.6-R1 masked video-prefix continuation PoC (2026-08-26)
|
||
|
||
- V3.6-R1 is a private experimental continuation transport only. The published V3.5.3 path remains `reference_context_v1`; no public widget, Production Node ID, saved-workflow schema, V3.4 contract, Sampling implementation, seed derivation, Plan, trim, Decode, Assembly, Seam, or Run Storage contract was changed.
|
||
- `masked_video_prefix_v1` copies the previous chunk's normalized 22-frame / 7T video tail into the next 42T video target, applies a full video mask with prefix `0` and generated region `1`, leaves the 235T audio target empty with an all-`1` mask, emits neither the old `minimax_refs` video context nor the Spectrum reference-context hint, and retains the existing 141-frame physical / 22-frame trim / 119-frame net contract.
|
||
- Core v0.34.0 honors the zero-mask prefix during model calls, but its raw final integrated float32 sampler output differed from the input prefix by up to `2.38418579101562e-07`. Continuum therefore restores only the protected target prefix after Sampling. The finalized physical latent is bit-exact (`torch.equal`, maximum absolute difference `0`, and SHA-256 match); generated video and audio regions are not overwritten.
|
||
- CPU/static acceptance is PASS: focused validation `8 passed`; canonical full pytest `500 passed` with only the known `pynvml` FutureWarning; changed-file `py_compile` and `git diff --check` pass.
|
||
- GPU Gate 1 and Gate 2 ran three fixed-seed A/B pairs through the normal ComfyUI_WAN API at `127.0.0.1:8188`: 2x5-second T2VA, 640x640, 20 steps, Balanced 22, Audio Continuity OFF, Run Storage OFF, no First/Last/Reference/Video Guide, Spectrum/Sol OFF, Seam OFF, and Sage Attention enabled globally. Every run preserved two physical samples (`124f` then `141f`, trim `22`, net `119`), matching derived seeds and chunk-1 video/audio hashes. All six outputs are finite 640x640 / 240-frame / 24 fps / 10.000-second MP4s.
|
||
- All masked runs report prefix exactness, zero/one video mask regions, all-one audio mask, empty second-group `minimax_refs`, and no old interop hint. API elapsed A/B pairs were `432.67/412.73`, `409.98/386.09`, and `408.62/386.56` seconds; these single ordered pairs are diagnostic observations, not an accepted speed claim.
|
||
- The initial 2x5 T2VA contact-sheet review found no clear pose snap, direction reversal, persistent artifact, or temporal collapse in either transport, but also no consistent visual superiority. The later 3x5 Strong Motion I2VA Stress Gate adds a positive signal without yet justifying a public/default replacement.
|
||
- Before the formal I2VA gate, diagnostics found that the private masked branch reapplied the original I2VA First Image on continuation chunks while `reference_context_v1` correctly removed it under `POLICY_REPLACE`. Those preliminary runs are invalid and excluded. The private branch now attaches the First Image only to chunk 1, so the accepted A/B differs only in continuation transport.
|
||
- Formal Stress Gate conditions were 3x5-second I2VA, 640x640, Balanced 22, three fixed seeds, Audio Continuity/Run Storage/Spectrum/Sol/Seam OFF, Prompt/CLIP cache HIT, `--disable-pinned-memory`, and alternating order `R->M`, `M->R`, `R->M`. All six normal runs are finite 640x640 / 360-frame / 24 fps / 15.000-second MP4s.
|
||
- Masked continuation was faster for all three seeds. API reductions were `6.185%`, `7.491%`, and `5.544%` (paired median `6.185%`). Continuation Sampling reductions were `9.471%`, `12.029%`, and `8.309%` (paired median `9.471%`; reference `466.424s`, masked `426.798s`). A separate CUDA-synchronized seed pair still reduced continuation Sampling from `464.97s` to `437.10s` (`5.995%`), proving that the signal is not only asynchronous host timing.
|
||
- Actual Core packed layouts explain the result: continuation `reference_context_v1` is `20,569` rows (`2,800` `ref_img` rows), while `masked_video_prefix_v1` is `17,769` rows with no `cond` or `ref_img` rows. This is `2,800` fewer rows / `13.61%` below the reference layout. Peak process RSS and observed device memory remained comparable; no memory regression was accepted or claimed.
|
||
- Every masked boundary retained the finalized 7T prefix bit-exact by SHA-256 and maximum difference `0`. Boundary contact sheets found no gross snap, direction reversal, persistent artifact, or collapse. At the second boundary, automated optical-flow jerk and absolute luma/saturation drift were lower for Masked in all three seeds; flow-magnitude continuity was mixed, so this is an improvement signal rather than proof of universal visual superiority.
|
||
- R1 Stress Gate is therefore **PASS**, while Production promotion remains **HOLD**. It opened the separately specified R2 Audio-prefix research gate; the later R2-0 CPU/static result is recorded above. Feathering, References, Terminal Merge, public UI, Run Storage integration, and GPU R2 acceptance remain unimplemented V3.6 research.
|
||
- Stress evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v36-r1-stress-gate-20260826\normal_fair` and `sync_fair`; `analysis.json` and `contact_sheets` contain the aggregate metrics and boundary frames. Invalid pre-fairness runs under the sibling `normal` directory are retained only as diagnostic history and must not be cited. Accepted Stress snapshots are `pre-v36-r1-stress-gate-source-accepted-20260826_174839`, `pre-v36-r1-stress-gate-wan-diagnostics-accepted-20260826_174839`, and `pre-v36-r1-stress-i2va-fairness-wan-runtime-20260826_182905`.
|
||
- Final validation after the Stress Gate is `500 passed`; changed-file `py_compile`, V3.5.3 runtime verification, `git diff --check`, source/runtime SHA-256 for four R1 Production files plus the diagnostic node, and empty API queue all pass. The ComfyUI_WAN backend intentionally remains running after acceptance.
|
||
- Reusable private A/B workflow: `D:\Codex\_test_workflows\ComfyUI-H3-Continuum\MiniMax_H3_Continuum_V36_R1_MaskedPrefix_AB.json` (22 nodes, 24 links, three annotated groups, SHA-256 `0964D55DDE1671E0BB0CA04FD3DD1F0AF82A4173FB9E6D9CE75900FC9B6981CE`). It is a simplified derivative of the user-supplied V3.5 workflow and requires the ComfyUI_WAN-only diagnostic node; it is not a public template.
|
||
|
||
## V3.5.3 maintenance hotfix (2026-08-26)
|
||
|
||
- V3.5.3 is a distribution-integrity maintenance hotfix over published V3.5.2. Sampling, Conditioning payloads, Terminal Merge, Assembly, Seam, Run Storage, Prompt/CLIP caching, Video Guide optimization, V3.4 compatibility, node IDs, and backend socket keys are unchanged.
|
||
- Repaired the public Conditioning Bridge workflow links `349`/`350` so the shared Width/Height controls target Sampler V3.5 input slots `16`/`17`; removed stale output link IDs `171`/`172`/`173` from the public V3.4, V3.4 Turbo, and V3.5 workflows.
|
||
- `validate_reference_prompts()` now preserves absolute public `<Picture N>` numbers for Hybrid First/Last + Reference warnings. This matches the previously accepted ComfyUI_W behavior and changes report text only.
|
||
- Re-encoded the unreferenced legacy `docs/images/v31b-auto-resume.png` without changing any displayed pixel. All 19 packaged PNG files pass strict Pillow verification.
|
||
- Added regressions for all eight public workflow graphs and the Hybrid warning numbers. Focused validation is `30 passed`; canonical full pytest is `495 passed`, and the source/runtime V3.5.3 verifier passes native PackedLayout, Fixed 3x5 planning, and complete node registration.
|
||
- The release Manifest now contains 173 files and verifies every recorded SHA-256, including the new public-workflow integrity regression.
|
||
- Production files `reference.py`, `version.py`, and `metadata.ini` are synchronized to ComfyUI_W with exact SHA-256 equality. ComfyUI_WAN is excluded because that secondary runtime is an incomplete older copy and must not be partially relabeled as V3.5.3.
|
||
- Accepted snapshots are `pre-v353-maintenance-hotfix-source-accepted-20260826_030732` (199 tracked files, revision `70cbbcd`) and `pre-v353-maintenance-hotfix-comfyui-w-accepted-20260826_030731` (14 present target files). The maintenance release was published as commit `67a0d80117bd8332d2768d093dff343001c87d60`.
|
||
|
||
## V3.5.2 final release gate (2026-08-26)
|
||
|
||
- V3.5.2 is the release-candidate baseline for the Stabilization & Optimization Update. It adds no generation mode and preserves all V3.5.x workflow contracts plus V3.4 Node IDs/backend socket keys.
|
||
- Phase 3-1 Prompt/CLIP cross-run cache and Phase 4-6A Video Guide memory optimization are the only adopted Production optimizations. Sampling, Conditioning payloads, Terminal Merge, Assembly, Seam, Run Storage, and standard ComfyUI Seed behavior remain unchanged.
|
||
- Final pre-release T2VA API smoke at 576x576 / 1x5 seconds / 20 steps confirmed MISS `encode_calls=1` and HIT `encode_calls=0`; Prompt/CLIP time changed from `8.454306s` to `0.000038s`. Both runs used derived Sampling seed `5684542245315818407` and produced identical decoded video MD5 `40e378d7c80e6bebd9d64b322c13e677` and audio PCM MD5 `4fb0a699227c4f834f7a52ba4d395ae8`.
|
||
- The accepted Video Guide GPU A/B remains the Phase 4-4 result: full source SHA-256, identity, 124-frame prefix, VAE/Qwen inputs, Sampling seed, decoded video MD5, and audio PCM MD5 all match exactly between old and optimized preprocessing.
|
||
- Final post-sync gate is PASS: `485 passed`, compileall exit `0`, V3.5.2 runtime verifier PASS, all 172 Manifest hashes valid, `git diff --check` exit `0`, required node registration complete, and the API queue empty. ComfyUI_W restarted with `H3 Continuum 3.5.2 loaded`; source, ComfyUI_W, and ComfyUI_WAN hashes match for `version.py`, `metadata.ini`, and the accepted `reference_video.py`.
|
||
- Public result image is `docs/images/v352-stabilization-optimization.png`, copied bit-exact from the user-supplied `D:\SD_PIC\Neo\Clip_91.png` (SHA-256 `0c1dfd5acff349b98ecdfb57ae50941f6da26a462dfd2aaa9b7c7b0a34abf304`). README text is the canonical environment description where the supplied image contains a different ComfyUI minor-version label.
|
||
- Exact release-preparation snapshot: `pre-v352-release-preparation-exact-20260826_022016`, revision `b8ab766eb364063f8af88781e7be70ee7286fba0`, 198 source files plus the four active runtime metadata files.
|
||
|
||
## V3.5.1 stabilization Phase 4-6A Video Guide memory optimization (2026-08-26)
|
||
|
||
- Phase 4 Feature Tax & Redundancy Audit is PASS. No evidence-backed Production speed optimization was selected. Driving Audio generated-audio Decode/Assembly and same-rate Audio hashing remain documented `SIMPLIFY` candidates below the 1% adoption threshold; Sampling, Session/State validation, Refine Context, Run Storage, Assembly Plan validation, and V3.4 remain unchanged.
|
||
- Phase 4-6A is formally PASS, including Phase 4-4 GPU acceptance. `reference_video.py` now normalizes, finite-checks, and hashes the complete Video Guide in eight-frame CPU float32 RGB chunks while retaining only the prefix used by the one-chunk H3 conditioning contract. Full source shape/dtype/raw-byte order and SHA-256 remain exact, so unused-tail changes and non-finite values still affect identity/validation.
|
||
- The primary ComfyUI_W runtime contained a pre-existing uncommitted Video Guide preprocess v3 that was newer than the repository v2. With explicit approval, its H3 `17k+5` frame alignment and final-frame padding behavior was integrated into the authoritative source before applying the memory optimization. The active v3 contract, source SHA-256, combined hash, frame count, and Run Storage identity are preserved by dedicated regressions.
|
||
- Deterministic 300x256x256 RGB / 124-frame-prefix comparison against the pre-change ComfyUI_W v3 produced the same `source_sha256` and complete contract. Preprocessing peak RSS above the already allocated input fell from `396.758 MiB` to `114.695 MiB`; retained Video Guide storage fell from `225.0 MiB` / 300 frames to `93.0 MiB` / 124 frames. Host time was `0.246672s` versus `0.257602s`, so this is accepted as a memory optimization rather than a speed claim.
|
||
- Dedicated Video Guide plus existing Driving Audio regressions are `17 passed`; canonical full pytest is `485 passed`. Source runtime verification, 171 Manifest hashes, syntax, and `git diff --check` pass. Only the known `pynvml` and inaccessible repository pytest-cache warnings remain.
|
||
- Phase 4-4 compared the pre-change v3 implementation and the optimized implementation through the normal ComfyUI API with the same 15-second / 360-frame 800x800 Video Guide, V3.5 one-chunk 5-second workflow, 576x576 output, 20 steps, seed input `46001`, and Hi-Res Fix disabled. Both resolved the guide to 640x640 / 124 frames, used one physical group and the same derived Sampling seed `3288707263902886519`, and assembled 124->120 frames at 24 fps / 5.000 seconds with 32 kHz stereo audio.
|
||
- The actual VHS-loaded source SHA-256 is `ea03aa0cb216213b3a7320f7fa262f226e8e9075d31f69bdb184eec7ea189f5d`; old/new combined identity, 124-frame prefix, 640x640 VAE input, and Qwen input are byte-identical. Decoded video MD5 is `53cd381aff184c2f8aa5910ab62be342` and decoded audio PCM MD5 is `3653c0ec89947a928075d4ba056ff624` on both paths. Representative frames show no gross artifact or collapse. Runtime was restored to the optimized file and its SHA-256 matches source (`a0266ab4434b8e9844219d2b6dd8a168d198dd8f142775be01a39b81dc196b91`).
|
||
- Accepted pre-change snapshot is `pre-phase4-6a-video-guide-memory-exact-20260826_012227` (6 exact source/runtime files, revision `b8ab766eb364063f8af88781e7be70ee7286fba0`). The standard full snapshot attempt stopped only on the known ACL-inaccessible `.pytest_cache`.
|
||
|
||
## V3.5.1 stabilization Phase 3-2 synchronized GPU rebaseline (2026-08-26)
|
||
|
||
- Phase 3-2 is formally PASS and is measurement-only. No Production node, schema, Sampling/Conditioning payload, Terminal Merge, Assembly, Seam, Run Storage, or V3.4 behavior changed. Diagnostics run from repository tools against the normal ComfyUI API and are not installed into `custom_nodes`.
|
||
- After one excluded warm-up and with the Phase 3-1 Prompt/CLIP cache already hot (`encode_calls=0`), the three-run API medians are `168.069s` for 1x5 T2VA at 576x576 and `379.765s` for 3x5 FL2VA Long Terminal Merge at 640x640. Synchronized in-process medians are respectively `166.986s` / `378.107s`, with Sampling `120.516s` / `295.216s`, Video VAE Decode `19.154s` / `51.634s`, Audio VAE Decode `0.466s` / `1.040s`, Assemble + Seam `0.130s` / `0.619s`, and MP4 save `2.583s` / `7.248s`.
|
||
- True whole-workflow PyTorch CUDA allocator peaks are `1.456 GiB` T2VA and `4.129 GiB` FL2VA. They are allocator observations, not total device VRAM usage and do not include every model-weight, ComfyUI staging, driver, or external allocation. Separate per-node peak runs confirm Sampling as the CUDA maximum; Video VAE Decode peaks are `0.290 GiB` / `0.319 GiB`. Whole-workflow Windows RSS/private peaks are about `29.15/45.53 GiB` T2VA and `30.82/46.18 GiB` FL2VA. These process peaks include Core/model staging and mapped/output pages and are attribution data, not Continuum-retained memory claims.
|
||
- T2VA external-attention API medians are: Sage `168.069s`; Sage + Sol-Attn `162.090s` (`-3.56%`); Sage + Spectrum `124.830s` (`-25.73%`); Sage + Sol-Attn + Spectrum `121.147s` (`-27.92%`). Sol's sparse Triton kernel was verified active with zero fallback after moving its writable cache to the test evidence root. Spectrum emitted its Continuum API acceptance marker. These routes may change floating-point results and are not bit-exact Continuum A/B claims.
|
||
- The fastest combined route also passed one measured FL2VA Long Terminal Merge run at `288.548s` API (`-24.02%` versus the Sage median): 3 logical / 2 physical groups (`[1]`, `[2,3]`), terminal seed/trim contract intact, 640x640, 360 frames, 24 fps, 15.000s, and audio present. All measured T2VA routes produced 576x576, 120 frames, 24 fps, 5.000s, and audio. Representative contact-sheet review found no gross line/plate artifact or collapse; this does not turn the external plugins into Continuum-owned quality guarantees.
|
||
- The new rebaseline leaves no evidence-backed Continuum Production optimization candidate for Phase 3-3. Sampling is dominant but contract-sensitive; Video/Audio VAE Decode and MP4 encoding are external Core/downstream work; Assemble + Seam is below the 1% adoption threshold. Phase 3-1 remains frozen rather than broadening its cache scope.
|
||
- Reusable diagnostics are `tools/p32_profile_server.py`, `tools/p32_api_runner.py`, and `tools/p32_gpu_profile_probe.py`. Accepted evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-2-20260825`. The accepted pre-change snapshot is `pre-stabilization-phase3-2-profile-accepted-20260825_220920` (479 files, revision `b8ab766eb364063f8af88781e7be70ee7286fba0`).
|
||
- Phase 3-2 focused validation is `13 passed`; canonical full pytest is `480 passed`. Diagnostic-file `py_compile`, all 170 Manifest hashes, source runtime verification, and `git diff --check` pass. Only the known `pynvml` FutureWarning remains.
|
||
|
||
## V3.5.1 stabilization Phase 3-1 Prompt/CLIP cross-run cache (2026-08-25)
|
||
|
||
- V3.5 now enables one private, bounded Prompt/CLIP conditioning cache. V3.4 public schema and execution remain unchanged. The cache belongs to the exact CLIP object, holds at most 16 LRU entries, and keys prompt text, exact resized First/Last image fingerprints, Last Frame presence, CLIP patch UUID, and CLIP layer.
|
||
- Reference Image, Reference Audio, Video Reference, Timeline Video, CLIP schedules, tokenizer options, and hook-modified CLIP states conservatively bypass the cache. Any cache error falls back to a normal encode and does not stop generation. Cached tensor payloads may be shared, but mutable metadata containers are copied for every consumer. CUDA-resident conditioning is never stored.
|
||
- Detailed Report records `hits`, `misses`, `bypasses`, and the actual Prompt/CLIP `encode_calls`. CPU regressions cover bit-exact payloads, mutable metadata isolation, every key invalidation, special-state bypass, CUDA-output rejection, fail-soft behavior, and 17th-entry LRU eviction.
|
||
- GPU/API 1x5 T2VA at 576x576 passed. MISS was `encode_calls=1`, Prompt/CLIP `5.898398s`; HIT was `encode_calls=0`, Prompt/CLIP `0.000028s`. Both same-seed outputs contain the same 120 decoded frame MD5 sequence and identical decoded audio PCM MD5. A separate base-seed-only change also remained a HIT with `encode_calls=0`, produced a different derived Sampling seed, and completed at 120 frames / 5.000s.
|
||
- GPU/API 3x5 FL2VA Long Terminal Merge at 640x640 passed with 3 logical / 2 physical groups (`[1]`, `[2,3]`). MISS was `encode_calls=4`, Prompt/CLIP `21.642687s`; HIT was `encode_calls=0`, Prompt/CLIP `0.007128s`. Both same-seed outputs contain the same 360 decoded frame MD5 sequence, identical decoded audio PCM MD5, and 15.000s duration. A separate base-seed-only change also retained four HITs / zero encode calls and the same physical grouping with newly derived physical seeds.
|
||
- No retained CUDA regression was observed: HIT completion CUDA allocated was lower than the adjacent MISS in both cases, and retained Refine Context GPU / observed GPU storage remained zero. The warm FL2VA HIT completed with about 21.3 MiB higher RSS than the adjacent MISS, consistent with the intentionally retained four-entry CPU conditioning cache and bounded by the 16-entry LRU.
|
||
- Focused CPU/static validation is `25 passed`; canonical full pytest is `467 passed`. Syntax, source/runtime verifier, source/runtime SHA-256 equality for the five changed Production files, and `git diff --check` pass. Accepted evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase3-1-20260825`.
|
||
|
||
## V3.5.1 stabilization Phase 2-2 Conditioning attribution (2026-08-25)
|
||
|
||
- Extended the existing V3.5 Detailed Report timing only; no node, socket, UI, Sampling, Conditioning payload, Terminal Merge, Assembly, Seam, or Run Storage contract changed. Basic/Off reports and every V3.4 path remain unchanged.
|
||
- Conditioning is now split into First/Last image VAE, Reference Image VAE, Reference Audio VAE, Driving Audio VAE, Video Reference VAE, Timeline Video VAE, and Prompt/CLIP. Unconnected paths report `n/a`, and counts distinguish an absent path from a measured near-zero path. Timing remains `time.perf_counter()` only with no CUDA synchronization or allocator mutation.
|
||
- CPU/static validation is `29 passed` focused and `453 passed` full. The first full run was blocked only by the existing Windows ACL on pytest's default temp root; the clean rerun with an explicit writable `--basetemp` passed. The remaining warnings are the known `pynvml` deprecation and inaccessible repository `.pytest_cache` cache-write warning.
|
||
- Accepted GPU/API 1x5 T2VA at 576x576 took 158.701s API / 135.970s Sampler host wall. Conditioning was 6.004392s, of which Prompt/CLIP was 6.004374s; identity video VAE was 0.000006s. A preceding warm-up independently measured 6.283131s Conditioning / 6.283109s Prompt/CLIP. A cached 3.398s replay is explicitly excluded.
|
||
- Accepted 3x5 FL2VA Long Terminal Merge at 640x640 took 410.727s API / 346.197s Sampler host wall and preserved 3 logical / 2 physical groups (`[1]`, `[2,3]`), 360 frames, 15.000s, and Exact Duration `362 -> 360`. Conditioning was 23.644903s: Prompt/CLIP 22.873708s (96.7%) and First/Last image VAE 0.771176s (3.3%). Sampling remained dominant at 322.335226s.
|
||
- No optimization is adopted. The measured Continuum bookkeeping remains negligible, and Prompt/CLIP time is required Core model encoding across the distinct physical prompts, including the Terminal Merge prompt. Reference-specific paths are instrumented but were not connected in these accepted runs, so their GPU cost remains unmeasured rather than assumed.
|
||
- Accepted evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-2-20260825`.
|
||
|
||
## V3.5.1 stabilization Phase 2-1 performance attribution (2026-08-25)
|
||
|
||
- Added measurement-only host timing to the existing V3.5 Detailed Report path. V3.4, Basic/Off reports, node schemas, UI, Sampling inputs/outputs, Terminal Merge, Assembly, Seam, and Run Storage contracts are unchanged.
|
||
- The report records total Preparation with Conditioning as an identified subset; each physical group's preparation, Sampling, CPU commit/finite validation, and Refine Context capture; plus Finalization and total host wall time. Terminal Merge is reported in physical order, so a 3-logical FL2VA run remains two measured groups: `[1]` and `[2,3]`.
|
||
- Timing uses `time.perf_counter()` only. It does not call `torch.cuda.synchronize()`, reset allocator peaks, clear caches, or change the normal asynchronous execution contract. Sampling values are host-observed elapsed time, not CUDA-event kernel timing.
|
||
- Core Video/Audio VAE Decode remains external and is not attributed to the Sampler. V3.5 Assembly continues to report its existing workflow elapsed time separately.
|
||
- CPU/static validation is `27 passed` focused and `451 passed` full with only the known `pynvml` FutureWarning. Runtime verification and source/runtime hashes pass.
|
||
- GPU/API measurement is PASS on the RTX 5060 Ti 16 GB system. Warm 1x5 T2VA at 576x576 took 143.888s API wall time; Sampler host wall was 134.736s, including 128.737s Sampling (95.5%), 5.963s Conditioning (4.4%), 0.013s CPU commit/finite validation, and 0.001s Refine Context capture. Output is 576x576, 120 frames, 5.000s, 32 kHz.
|
||
- The 3x5 FL2VA Long Terminal Merge run took 396.477s API wall time and retained 3 logical / 2 physical groups (`[1]`, `[2,3]`). Sampler host wall was 348.047s: 323.835s Sampling (93.0%), 24.003s Conditioning (6.9%), 0.050s CPU commit/finite validation, and 0.010s Refine Context capture. Output is 640x640, 360 frames, 15.000s, 32 kHz.
|
||
- The 1x5 Hi-Res Fix 1.2x run took 330.005s and produced 704x704, 120 frames, 5.000s, 32 kHz. Against the warm non-Hi-Res baseline, the integrated Resize/Second Pass/Decode path added about 186.1s, 0.49 GiB peak Working Set, and 1.34 GiB peak Private Bytes. Global observed VRAM usage was about 15.4-15.6 GiB in all three runs and is not a per-node allocator measurement.
|
||
- No optimization has been selected or implemented. Sampling dominates and the measured CPU commit, finite validation, Refine Context, and group preparation paths are rejected as optimization targets. Conditioning exceeds the 1% investigation threshold but must be split into CLIP/VAE/reference subphases before any change is justified.
|
||
- Accepted evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\stabilization-phase2-20260825`. The first 1x5 attempt reached an older already-running ComfyUI process and is retained only as excluded diagnostic evidence; all accepted runs used the restarted updated listener.
|
||
|
||
## V3.5.1 stabilization Phase 0 CI portability candidate (2026-08-25)
|
||
|
||
- V3.5.1 is feature-frozen for stabilization. Speculative performance changes and broad refactors are excluded; only measured bottlenecks may proceed after the remaining acceptance work closes.
|
||
- The P2-5 process-memory regression is platform-aware without changing the probe or any Production path: RSS is required everywhere, USS is checked when the platform exposes it, and Windows private commit is tested only on Windows.
|
||
- Local focused validation is `4 passed`; full Windows validation is `453 passed` with only the known `pynvml` FutureWarning. Syntax and `git diff --check` pass. Remote GitHub Actions remains to be confirmed before publication or updating the public validation count.
|
||
|
||
## V3.5.1 Issue #9 Conditioning Bridge candidate (2026-08-24)
|
||
|
||
- Added `H3ContinuumConditioningBridgeV35` / `H3 Continuum Conditioning Bridge V3.5` under `MiniMax H3/Continuum/Advanced`.
|
||
- The Bridge accepts externally processed physical-group video LATENTs plus the matching Assembly Plan and required V3.5 refine context. It returns one paired MODEL and one complete ComfyUI CONDITIONING object per physical group, plus a target-geometry Assembly Plan and status.
|
||
- The fixed list contract is `OUTPUT_IS_LIST = (True, True, False, False)` and `len(group_models) == len(conditioning) == physical group count`. CONDITIONING entries are appended as one inner object and are never flattened into the outer physical-group list.
|
||
- Second Pass and Bridge share the same context validation, target adaptation, prompt fallback, MODEL clone, and geometry update helper. Built-in Second Pass still samples groups sequentially and retains responsibility for Nested AV, video noise, Audio Lock, and audio passthrough; the Bridge deliberately leaves AV LATENT construction, noise, Audio Lock, and external sampling to the workflow.
|
||
- CPU coverage includes the 3-logical / 2-physical Terminal Merge shape (`[1] / 37T`, `[2,3] / 77T`), MODEL/CONDITIONING alignment, non-flattening, schema/registration, and existing Reference Image/Audio/context adaptation regressions. Focused validation is `33 passed`; full source pytest is `442 passed` with only the known `pynvml` FutureWarning.
|
||
- Only `nodes.py`, `v3/second_pass.py`, `v3/conditioning_bridge_nodes.py`, and `tools/verify_runtime.py` were synchronized to ComfyUI_W; every source/runtime SHA-256 matches. Runtime registration, native PackedLayout, and Fixed 3x5 prompt planning pass with the new Bridge registered as a non-legacy node.
|
||
- GPU acceptance through Basic Guider and an external sampler passed for three `1 x 5s` runs: LBH `1.2x` with `simple / 4 steps / denoise 0.3` produced 704x704 twice, and LBH `1.5x` with `simple / 10 steps / denoise 0.3` produced 864x864. Every run kept one MODEL and one complete CONDITIONING for one physical group, completed external sampling/Core decode/Assembly, and produced finite 120-frame/5.000-second video with 32 kHz stereo audio. The existing Issue #10 Reference Audio working-tree changes were preserved without reset or reorganization.
|
||
|
||
## V3.5.1 Issue #10 Reference Audio candidate (2026-08-24)
|
||
|
||
- Target version is V3.5.1. The optional `reference_audio_1` and `reference_audio_vae` sockets remain permanently present in their Python-defined input order and are displayed as `Reference Audio (Optional)` and `Reference Audio VAE (Optional)`.
|
||
- The earlier frontend-only `Reference Audio Inputs` dropdown and all dynamic socket add/remove behavior were removed by the compatibility hotfix. Workflow serialization now contains only the Python-defined widgets, preserving positional widget-value alignment across save/reload.
|
||
- Related display labels are now `Video Guide Frames`, `Driving Audio`, and `Driving Audio VAE`. Input keys, order, types, Node IDs, and runtime contracts remain unchanged.
|
||
- This exposes the existing Core-compatible standalone Reference Audio 1 path; it adds no new audio encoding, strength control, model allowlist, source-audio copy, or lip-sync guarantee.
|
||
- V3.4 input keys/order/types, Node ID, sockets, and behavior remain unchanged; only the human-facing labels/tooltips are clarified. Its internal `run()` accepts the two values only so V3.5 can forward them to the existing Production path.
|
||
- Reference Audio conditions generated video/audio and leaves generated audio as final. Driving Audio remains the separate timeline-guide path whose original source is selected as final output.
|
||
- CPU integration confirms Reference Audio survives every normal physical group and the FL2VA Terminal Merge group (`logical [2,3]`, `paired_timeline_v1`) in `refine_context`. Run Storage ignores an unused Reference Audio VAE, separates changed audio contracts, and restores the disconnected revision identity.
|
||
- Final canonical validation after the conditional UI change is `450 passed` with only the known `pynvml` FutureWarning. JavaScript syntax, `git diff --check`, ComfyUI_W registration/native PackedLayout/Fixed 3x5 verification, and source/runtime SHA-256 equality for synchronized implementation files pass.
|
||
- The earlier conditional-visibility UI acceptance is superseded: a saved/reloaded workflow exposed positional widget corruption under the current frontend. The permanent-socket hotfix removes that frontend mutation path without changing Reference Audio execution.
|
||
- The CPU-validated public feature files were previously published from the isolated clone as `a704aa961fc2ba0c19aef6f2ecfdcb46fa9384b2`. GPU acceptance is now complete and package metadata plus public documentation are being formalized as V3.5.1.
|
||
- The pre-hotfix V3.5.1 release baseline was `450 passed`. Hotfix validation is `452 passed` with explicit workflow save/reload widget-alignment coverage and only the known `pynvml` FutureWarning. JavaScript syntax and `git diff --check` pass. The two frontend files are synchronized to ComfyUI_W by SHA-256; live save/reload acceptance remains pending.
|
||
- The public LBH/Conditioning Bridge connection example is `examples/workflows/MiniMax_H3_Continuum_V351_LBH_Conditioning_Bridge.json`; its README chart is `docs/images/v351-lbh-conditioning-bridge-flow.svg`. It is a separate example and does not replace or modify the recommended V3.5 template.
|
||
|
||
## V3.5.1 standard ComfyUI Seed behavior (2026-08-25)
|
||
|
||
- The experimental Last Queued Seed frontend override was withdrawn after repeated real-browser failures. Static callback and simulated-event tests did not reproduce the active ComfyUI Frontend lifecycle reliably.
|
||
- `H3 Continuum Sampler V3.5` now uses the same native `control after generate` behavior as ComfyUI Core. Continuum does not wrap seed callbacks, intercept Randomize/Fixed transitions, monitor widgets, or rewrite submitted seed values.
|
||
- This removal is limited to the Last Queued Seed experiment. V3.4/V3.5 node schemas, Reference Audio UI, compact UI settings, Sampling, Second Pass, Terminal Merge, Assembly, Seam, Run Storage, and all other frontend behavior remain unchanged.
|
||
- Focused unaffected-UI validation is `8 passed`; full source pytest is `450 passed` with only the known `pynvml` FutureWarning.
|
||
|
||
## V3.5.1 Hi-Res Fix Core canvas alignment (2026-08-24)
|
||
|
||
- A real `576x576`, `scale_by=1.2` run exposed an invalid `43x43` H3 video/conditioning latent. Core MiniMax H3 patchifies spatial latents in `2x2` blocks and exposes pixel dimensions in 32-pixel steps, so the odd latent failed in Core with `24 * 43 * 43 = 44376` elements.
|
||
- Manual Hi-Res Fix targets now use ComfyUI Core's `CANVAS_MULTIPLE` and preserve-or-enlarge each axis on that grid. The failing case resolves to `704x704` / `44x44`; a `1024x576` source at `1.2x` resolves to `1216x704` / `76x44`. The Pixel/VAE, conditioning adaptation, and Second Pass paths remain unchanged.
|
||
- Focused Hi-Res/Second Pass validation is `48 passed`; full source pytest is `448 passed`. `v3/hires_fix_nodes.py` is synchronized to ComfyUI_W by SHA-256, and runtime registration/native PackedLayout/Fixed 3x5 verification passes. GPU acceptance after a ComfyUI process restart remains pending.
|
||
|
||
## V3.5 recommended template workflow (2026-08-24)
|
||
|
||
- `examples/workflows/MiniMax_H3_Continuum_V35.json` is the recommended V3.5 starting workflow.
|
||
- Hi-Res Fix is disabled by default and `H3 Continuum Assemble + Seam V3.5` uses Auto backend selection.
|
||
- First Frame, Last Frame, and Reference Images 1-3 share one 0.3 MP control. Video Guide Frames remains independently sized.
|
||
- Reference Images 1-3 and Video Guide Frames are routed to their matching V3.5 sampler sockets; `refine_context` and `video_vae` are connected to the integrated Hi-Res Fix path.
|
||
- V3.4 standard and Turbo templates remain available for saved-workflow compatibility.
|
||
|
||
## Baseline
|
||
|
||
- Published baseline: V3.5.3 maintenance hotfix (`67a0d80117bd8332d2768d093dff343001c87d60`)
|
||
- Authoritative working repository: `D:\Codex\_git_push_work\ComfyUI-H3-Continuum`
|
||
- Primary tested runtime: `D:\StabilityMatrix\Data\Packages\ComfyUI_W\custom_nodes\ComfyUI-H3-Continuum`
|
||
- Secondary runtime: `D:\StabilityMatrix\Data\Packages\ComfyUI_WAN\custom_nodes\ComfyUI-H3-Continuum`
|
||
- Old checkout to keep synchronized with published `main`: `D:\Codex\ComfyUI-H3-Continuum`
|
||
|
||
## Active direction
|
||
|
||
- Driving Audio preserves the supplied audio as final output while conditioning every chunk.
|
||
- Video Guide Frames is persistent across chunks and supports selectable guide sizing.
|
||
- Prompt syntax is permissive: invalid or empty input warns and falls back instead of blocking generation.
|
||
- Unknown models, wrappers, merged models, and upstream nodes are not rejected merely because they are unknown.
|
||
- Legacy `strict_compatibility` remains loadable only for workflow compatibility and must not enforce a stop.
|
||
- Keep the compact V3.4 interface and preserve all public sockets and node identifiers unless a UI/API change is explicitly approved.
|
||
|
||
## Current working tree
|
||
|
||
The authoritative repository may contain active, uncommitted Hybrid reference work. Preserve it. Never reset, overwrite, or publish it without explicit instruction.
|
||
|
||
## Standard procedure
|
||
|
||
1. Read `AGENTS.md` and this file.
|
||
2. Inspect `git status -sb` and confirm the authoritative source/runtime.
|
||
3. Run `tools\snapshot.ps1` before edits.
|
||
4. Make the smallest relevant change.
|
||
5. Run `tools\validate.ps1` only when validation is requested or publication is authorized.
|
||
6. Append the result and remaining gaps to `WORKLOG.md`.
|
||
7. Commit or push only when explicitly requested.
|
||
|
||
## Validation boundary
|
||
|
||
Static checks do not replace a real GPU generation. Record workflow, prompt, media, model/LoRA, sampler, steps, dimensions, chunks, elapsed time, and observed result for GPU tests.
|
||
|
||
## Active FLF implementation
|
||
|
||
- Hybrid FLF Chunk Contract v1 is implemented locally in `v2/sequence.py`.
|
||
- It replaces the previous single global sample/split strategy with a 0.4 MP full-sequence guide, intermediate boundary anchors, and separately sampled FL2VA chunks.
|
||
- Reference Images remain persistent in the guide and every output chunk.
|
||
- Real GPU generation is still required before this path can be considered stable.
|
||
|
||
## FLF GPU regression baseline (2026-08-21)
|
||
|
||
- PASS `MiniMax_H3_Continuum_V31_00029_.mp4`: pure FLF skateboard sequence; visual continuity accepted.
|
||
- PASS `MiniMax_H3_Continuum_V31_00031_.mp4`: Hybrid model plus Reference Image; visual result accepted.
|
||
- PASS `MiniMax_H3_Continuum_V31_00032_.mp4`: standard Ref2VA model; visual result accepted.
|
||
- FAIL `MiniMax_H3_Continuum_V31_00034_.mp4`: Hybrid FLF+Reference; visible jumps near 5 s and 9 s.
|
||
- FAIL `MiniMax_H3_Continuum_V31_00035_.mp4`: Ref2VA model plus FL2V Turbo 8-step LoRA; two visible flashes/jumps.
|
||
- Current FLF candidate is `context_bound_terminal_last_only_v2`: the final chunk keeps its 22-frame continuation context and the single Last Frame anchor retained by `prepare_conditioning()`, without adding a duplicate context-tail boundary keyframe.
|
||
- GPU verification of this candidate is pending; it is not yet a stable baseline.
|
||
- `context_bound_terminal_bridge_v2` is withdrawn because its synthetic intermediate latent correlated with the later terminal flash regression.
|
||
- Preserve Hybrid/Reference support, Run Storage identity, UI, audio paths, and permissive prompt handling when repairing FLF.
|
||
|
||
## FLF GPU regression update: 00036 (2026-08-21)
|
||
|
||
- FAIL `MiniMax_H3_Continuum_V31_00036_.mp4`.
|
||
- Workflow: Hybrid B20-49, FL2V Turbo 8-step LoRA, 8 steps, Euler/simple, Spectrum Off, 2 x 5 s, 640x640, Video Reference Efficient 0.4 MP, randomized seed.
|
||
- The 5 s chunk boundary was visually stable, but a bright ghost/flash and discontinuous disappearance of the reference fox occurred near 8 s.
|
||
- Project acceptance requires whole-sequence continuity; a clean chunk seam alone is insufficient. This run is not a partial pass.
|
||
- Restoring `context_bound_terminal_v1` did not resolve the terminal Hybrid FLF+Reference discontinuity and is not accepted as the final repair.
|
||
|
||
## FLF GPU regression update: 00037 (2026-08-21)
|
||
|
||
- FAIL `MiniMax_H3_Continuum_V31_00037_.mp4`.
|
||
- Workflow: official FL2VA pruned INT8 ConvRot model, FL2V Turbo 8-step LoRA, 8 steps, Euler/simple, Spectrum Off, 2 x 5 s, 640x640, one reference image, Video Reference Efficient 0.4 MP, randomized seed.
|
||
- The 5 s chunk boundary remained visually stable, but the reference fox became translucent and disappeared with a bright ghost/flash near 7.5-8 s before the Last Frame composition.
|
||
- This reproduces the terminal discontinuity seen in 00036 and is a whole-sequence failure.
|
||
- Do not roll back merely to terminal bridge v2: 00034 and 00035 already failed there. Restore only an exact baseline associated with the accepted 00029/00031 behavior.
|
||
|
||
## FLF GPU regression update: 00038 and rollback (2026-08-21)
|
||
|
||
- FAIL `MiniMax_H3_Continuum_V31_00038_.mp4` by user visual evaluation.
|
||
- Embedded workflow: Ref2VA pruned INT8 ConvRot model, FL2V Turbo 8-step LoRA, Euler/simple, 8 steps, Spectrum Off, 2 x 5 s, randomized seed, Video Reference Efficient 0.4 MP.
|
||
- The initial rollback after 00038 temporarily returned to bridge v2, but the user clarified that the required target is the exact code state used for PASS 00031.
|
||
- `00031` completed at 2026-08-21 02:32:59; `pre-flf-terminal-bridge-v2-repo-20260821_023553` was captured immediately afterward and before bridge v2 was introduced.
|
||
- Current FLF strategy is restored exactly from that snapshot: `context_bound_terminal_v1`. Bridge v2 is withdrawn and is not active.
|
||
|
||
## M1 Hybrid diagnostics (2026-08-21)
|
||
|
||
- Canonical implementation source: ComfyUI_W runtime.
|
||
- Runtime pre-change snapshot: D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-m1-comfyui-w-canonical-20260821_172324.
|
||
- Added display-only Hybrid conditioning labels; the stored/runtime conditioning mode remains unchanged.
|
||
- Hybrid reference warnings now use the actual global Qwen <Picture N> numbering.
|
||
- Added README numbering guidance and focused CPU-only regression tests.
|
||
- No Sampling, Conditioning payload, Run Storage contract, UI socket, or model-selection behavior changed.
|
||
- Tests were added but not executed in this implementation step.
|
||
|
||
## Terminal Merge Physical Seed v2 (2026-08-22 00:49:57)
|
||
- Terminal Merge uses one physical seed per merged pair; an initial 2x5s pair uses base_seed directly when reroll is inactive.
|
||
- Later terminal pairs derive once from the zero-based pair-start index; both logical entries store that same physical seed.
|
||
- Run Storage distinguishes the scoped terminal_merged_10s_seed_v2 / physical_window_seed_v2 execution semantics without changing the global schema version.
|
||
|
||
|
||
## Terminal Merge diagnostic harness (2026-08-22)
|
||
|
||
- Production generation semantics remain unchanged; diagnostics live under tests only.
|
||
- Diagnostic stages are ordered A Pre-Sampler, B First MODEL Call, C sampled physical AV, D split/recombine, E decode/assembly.
|
||
- Tensor identity uses shape, dtype, and SHA-256 of contiguous CPU raw bytes; numerical mismatch also records max/mean absolute difference and allclose.
|
||
- Terminal split/recombine requires bit-exact equality.
|
||
- Real Core vs Continuum GPU capture remains pending; this unit-test phase does not load checkpoints.
|
||
|
||
## GPU Core/Continuum diagnostic runner (2026-08-22)
|
||
- Source: `tools/gpu_diagnostic_node/`
|
||
- Installer: `tools/install_gpu_diagnostic_node.ps1`
|
||
- Runtime: `D:\StabilityMatrix\Data\Packages\ComfyUI_W\custom_nodes\H3_Continuum_GPU_Diagnostics`
|
||
- Node: `H3 Core vs Continuum GPU Diagnostic`
|
||
- Production Continuum files and generation semantics are not modified by this diagnostic plugin.
|
||
- Captures A: pre-sampler, B: first MODEL call, C: physical sampled AV latent, D: terminal split/recombine, E1: decode comparison, E2: common 243-to-240 adjustment.
|
||
- Artifacts: JSON manifest plus safetensors under the configured ComfyUI output diagnostics directory.
|
||
- Pre-change snapshot: `D:\Codex\_snapshots\ComfyUI-H3-Continuum\pre-gpu-core-continuum-diagnostic-runner-20260822_022625` (138/138 files verified).
|
||
- GPU comparison has not yet been executed.
|
||
|
||
## Long Terminal Merge for distinct timeline sections (2026-08-22)
|
||
|
||
- For FL2VA with 3 or more 5-second logical chunks, the final two chunks may now share one physical Terminal Merge sample even when their prompt sections differ.
|
||
- Distinct final prompts are rebased to a local physical timeline (`[0-5s]` and `[5-10s]`); identical prompts retain the accepted 2x5 Core-equivalent prompt path unchanged.
|
||
- Expected 3x5 execution: 3 logical chunks, 2 physical sampling passes, and 2 physical VAE decode groups; the terminal group is 260 frames with a 22-frame decoded prefix trim and 238 retained frames.
|
||
- Timeline Video remains excluded from Terminal Merge. UI, public sockets, Conditioning payload construction, Physical Seed v2, Assembly, and Seam semantics are otherwise unchanged.
|
||
- GPU acceptance passed for `MiniMax_H3_Continuum_V31_00008_.mp4`: 3 logical chunks, 2 physical sampling passes, decoded groups of 124 and 238 retained frames, 362-to-360 exact-duration adjustment, and 15.000-second output.
|
||
- Core comparison `MiniMax_H3_00005_.mp4` showed the same short terminal still period. This is accepted as normal FL2VA Last Frame convergence, not a Continuum-specific defect.
|
||
- Long Terminal Merge is formally adopted. Do not retune Conditioning, Physical Seed v2, physical decode grouping, or terminal prompt pairing without new contrary regression evidence.
|
||
|
||
## V3.5 Hi-res Fix and Second Pass (2026-08-23)
|
||
|
||
- The original V3.5 Main candidate was first-pass physical video latents -> lightweight Core spatial resize -> low-sigma Second Pass -> Core Decode -> Assemble + Seam. The 2026-08-24 GPU isolation below rejects plain latent interpolation as the general Main backend; the Advanced Second Pass bridge remains valid.
|
||
- E2-8A and E2-8B1 established additive `assembly_plan.second_pass_contract` metadata and static physical-group validation without changing V3.4 generation semantics.
|
||
- E2-8B2 GPU PoC is accepted: LBH 3D Upscaler preserved H3 24-channel physical groups, including the 3x5 FL2VA 124f + 260f Terminal Merge layout.
|
||
- E2-8C adds an experimental T2VA-only physical-group Second Pass node. It reads prompts from the plan, uses workflow-supplied SIGMAS, derives one refine seed per physical group, adopts refined video, and returns first-pass audio LATENT objects unchanged.
|
||
- `H3ContinuumLatentResizeV35` is the standalone lightweight Utility. It delegates to ComfyUI Core `common_upscale`, exposes the same nearest-exact/bilinear/area/bicubic/bislerp methods and a scale multiplier, preserves physical-group order and B/C/T, and changes only latent H/W. It is not an accepted general-purpose Main Hi-res backend.
|
||
- The final V3.5 node layout keeps three distinct roles: `H3ContinuumHiResFixV35` is the normal one-node Main path, `H3ContinuumSecondPassV35` is the Advanced entry for LBH or other externally processed latents, and `H3ContinuumLatentResizeV35` remains a standalone Utility.
|
||
- Hi-Res Fix ON performs an H3 Video VAE Decode -> CPU pixel resize -> same H3 Video VAE Encode for each physical group, then calls the existing context-aware physical-group Second Pass helper. Hi-Res Fix OFF returns the exact first-pass video/audio LATENT lists and original Assembly Plan object; all sampling and VAE inputs remain lazy and are not requested by the node.
|
||
- Second Pass itself has no enable switch and no duplicate first-pass input. Its public display name is `H3 Continuum Second Pass V3.5`; the Utility display name is `H3 Continuum Latent Resize V3.5`.
|
||
- E2-8C GPU acceptance is formally PASS for 1x5 T2VA, 3x5 T2VA, and 3x5 FL2VA Long Terminal Merge.
|
||
- The accepted FL2VA run preserved two physical groups as `37T + 77T`: group 1 mapped logical `[1]`, and group 2 mapped logical `[2,3]` with `paired_timeline_v1`. Each physical group was sampled exactly once by Second Pass.
|
||
- Final output was 768x768, 360 frames at 24 fps, and 15.000 seconds. First-pass audio LATENT passthrough, Core Video/Audio Decode, and Assemble succeeded; the 5-second and 10-second boundaries had no major discontinuity.
|
||
- The final approximately 0.5-second low-motion interval is accepted as natural Last Frame convergence, not temporal collapse.
|
||
- V3.4 Sampling, Terminal Merge, Assembly, Seam, and Run Storage were not changed by E2-8C.
|
||
- First/Last, Reference, Hybrid, Run Storage, and embedded LBH integration remain outside E2-8C.
|
||
- The current V3.5 working tree uses a Second Pass-only direct-sampling contract: video receives seed-derived random noise with mask 1, while first-pass audio receives zero noise with mask 0. Core applies workflow SIGMAS once; no external AddNoise, inverse noise scaling, or `audio / (1-sigma0)` compensation is used.
|
||
- This path is isolated in `v3/refine_sampling.py`; V3.4 `v2/sampling.py::sample_chunk()` remains unchanged. The Assembly Plan records execution as `t2va_physical_groups_audio_locked_v2` and audio sampling as `zero_noise_mask_locked`.
|
||
- CPU validation is `277 passed`. Core `reshape_mask` also expands the minimal video `[B,1,1,1,1]` and audio `[B,1,1,1]` masks correctly for real H3 stream shapes.
|
||
- The audio-locked v2 path is now GPU-confirmed for direct no-resize Second Pass: `V35_NoResize_SecondPass_Diagnostic_00001_.mp4` used the cached 576x576 first-pass latent, `res_multistep`, `simple` 10 steps / denoise 0.35, and refine seed 42. Its report confirmed video random noise/mask 1, audio zero noise/mask 0, one 37T physical group, and one sampling pass.
|
||
- The direct no-resize result is visually clean across the full 120-frame / 5.000-second output. The cyan/orange/white line and plate artifacts remain in the otherwise matched one-node Hi-Res Fix run `E2-8B2_01_1x5_T2VA_LBH3D_00017_.mp4`, which used bilinear 2x from 576x576 to 1152x1152.
|
||
- The fixed-condition GPU isolation therefore rejects random audio re-noise as the primary artifact cause and validates the basic direct Second Pass re-noise path. The remaining cause is scoped to the 2x spatial-resize/high-resolution re-entry path; this result alone does not yet distinguish interpolation method from scale/resolution.
|
||
- ComfyUI_W registration, native PackedLayout, and Fixed 3x5 prompt-plan verification passed before this isolation. The one-node Hi-Res Fix wrapper executes successfully on GPU, but its bilinear 2x path is visually rejected and is not an accepted standard Hi-res Fix path.
|
||
- Resolution/method isolation used the same cached first-pass latent, `res_multistep`, `simple` 10 steps / denoise 0.35, and refine seed 42 throughout. Both nearest-exact and bilinear were clean at 576x576 -> 768x768 (`36x36 -> 48x48` latent), while bilinear failed at 1024x1024 (`64x64`) and both bilinear and nearest-exact failed at 1152x1152 (`72x72`).
|
||
- This rejects bilinear-specific channel mixing as the primary cause within the direct interpolated-latent path. The tested safe square target for that legacy path is 768x768; its exact failure threshold between 768 and 1024 was not mapped. ComfyUI Core's H3 implementation independently defines `BASE_SHORT_EDGE = 768` and `MAX_PIXELS = 768 * 1344`, consistent with the observed direct-re-entry boundary. These limits do not describe the later Pixel/VAE roundtrip result.
|
||
- The current `H3ContinuumHiResFixV35` defaults to pixel Lanczos and `scale_by=2.0`; zero retains the H3 native-canvas Auto option and positive values retain manual multiplier semantics. Enabled execution requires the lazy `video_vae` input and never falls back to direct latent interpolation. Disabled execution does not evaluate the VAE or sampling inputs.
|
||
- Manual mode is not clamped or stopped when it exceeds the Core native canvas; the status reports the target as informational and execution continues through the Pixel/VAE backend. No V3.4 constraint or independent target policy was added.
|
||
- `V35_NativeAuto_576to768_Acceptance_00001_.mp4` remains historical evidence for the former direct interpolated-latent backend only. Main Hi-Res Fix no longer uses that backend; the explicit `H3ContinuumLatentResizeV35` Utility remains available separately.
|
||
- The earlier recommendation to move all 2x delivery to a post-decode video upscaler is superseded for Main Hi-res Fix by the validated Pixel/VAE roundtrip in the next section. Issue #8's external-latent Second Pass bridge remains valid independently.
|
||
|
||
## V3.5 context-aware Hi-res isolation (2026-08-24)
|
||
|
||
- Added `H3ContinuumSamplerV35` without changing the V3.4 public contract. Its first five outputs are byte-for-byte ordered like V3.4 (`video_latents`, `audio_latents`, `assembly_plan`, `status`, `driving_audio`); output six is runtime-only `H3_CONTINUUM_REFINE_CONTEXT`.
|
||
- The context captures the exact pre-sampling conditioning once per physical group, including Terminal Merge logical mapping, physical clip index, context frames, prepared First/Last images, and source H3 latent geometry. It is never stored in Assembly Plan, Session, logical entries, or Run Storage.
|
||
- A complete context is adapted to target H/W by re-encoding First/Last keyframes with the connected Video VAE and resizing only marked Continuum video-context latents. Ordinary MiniMax references, text conditioning, and all audio conditioning remain unchanged. Incomplete Run Storage capture falls back to the legacy prompt-only path with a warning; a corrupt typed context is a contract error.
|
||
- Second Pass now creates one call-local Continuum MODEL clone per physical group. The audio-lock contract remains video random noise/mask 1, audio zero noise/mask 0, one workflow SIGMAS application, one refine seed and one sampling call per physical group, and bit-exact first-pass audio LATENT passthrough.
|
||
- CPU/static validation after integration is `293 passed` with only the known `pynvml` FutureWarning. Runtime registration exposes `H3 Continuum Sampler V3.5` with six outputs. V3.4 Sampling, Terminal Merge, Assembly, Seam, Run Storage, and its five-output UI contract remain unchanged.
|
||
- GPU A/B `0adff3b7-1552-47e9-a167-7b6e4010b0a5` used the fixed FL2VA 1x5 condition, 576x576 -> 864x864 bilinear 1.5x, `res_multistep`, `simple` 10 steps / denoise 0.35, and refine seed 42. The report confirmed `conditioning_source=refine_context`, two keyframes re-encoded at target H/W, shape `(1, 24, 37, 54, 54)`, and first-pass audio passthrough. Output `V35_ContextAware_FL2VA_576to864_d035_00001_.mp4` still contains the persistent cyan/orange/white line and plate artifacts. Context mismatch is therefore not the primary cause.
|
||
- Resize-only GPU isolation `75249bbe-306a-45f9-a139-a79c0b17e536` decoded a plain bilinear 576x576 -> 864x864 latent as `V35_ResizeOnly_FL2VA_576to864_bilinear_00001_.mp4`. It removed the colored lines but produced severe person/background ghosting, confirming that channel-wise interpolation creates an invalid H3 video-latent representation even before refinement.
|
||
- Pixel/VAE roundtrip GPU isolation `2a03a1d4-2355-4d5a-a237-0ec3e0c09761` used First Pass Decode -> pixel Lanczos 864x864 -> H3 Video VAE Encode -> the same context-aware denoise 0.35 Second Pass. `V35_PixelRoundtrip_FL2VA_576to864_d035_00001_.mp4` completed at 864x864, 120 frames, 24 fps, 5.000 seconds with no colored-line artifacts or resize ghosting while retaining the orange jacket, subject identity, motion, and terminal composition.
|
||
- Pixel/VAE roundtrip 2x GPU acceptance `de11e383-8b3d-4d8a-806f-1ac3550091f1` used the same cached First Pass, pixel Lanczos 1152x1152, H3 Video VAE Encode, context-aware `res_multistep` / `simple` 10-step denoise 0.35 Second Pass, and refine seed 42. It completed without OOM on the 16 GiB GPU as `(1, 24, 37, 72, 72)`, one physical group / one sampling pass, two target keyframes VAE re-encoded, first-pass audio passthrough, Core Decode/Audio Decode, and Assemble. `V35_PixelRoundtrip_FL2VA_576to1152_d035_00001_.mp4` is 1152x1152, 120 frames, 24 fps, and 5.000 seconds; twelve representative time points are visually clean with no colored lines/plates, resize ghosting, or temporal collapse, while subject identity, orange clothing, motion, and composition remain coherent.
|
||
- The evidence-backed Main backend is implemented as a sequential Pixel/VAE roundtrip per physical group followed by the existing context-aware Second Pass. Pixel resize uses bounded 8-frame CPU batches; VAE Encode matches Core's 4D IMAGE contract by encoding each Continuum batch as `[F,H,W,C]` and then restoring B. Plain interpolated latent is never used as a Main fallback. `H3ContinuumSecondPassV35` remains the Advanced external-LBH/learned-upscaler bridge, and `H3ContinuumLatentResizeV35` remains an explicit Utility.
|
||
- CPU/static validation after the Main backend implementation is `308 passed`; ComfyUI_W runtime registration, native PackedLayout, Fixed 3x5 prompt-plan verification, `py_compile`, and `git diff --check` pass. The only test-environment messages are the known pytest cache permission warning and `pynvml` FutureWarning.
|
||
- Integrated Main GPU acceptance prompt `c5dbe55c-c846-47e8-9f49-8bb167887236` used FL2VA 1x5, 576x576 -> 1152x1152, pixel Lanczos, context-aware `res_multistep` / `simple` 10-step denoise 0.35 Second Pass, and refine seed input 42. It completed in 14:18 as one physical group / one sampling pass with `(1, 24, 37, 72, 72)`, two target keyframes re-encoded, and first-pass audio LATENT passthrough.
|
||
- `V35_MainSafePixelVAE_Core4D_FL2VA_576to1152_d035_00001_.mp4` is 1152x1152, 120 frames, 24 fps, and 5.000 seconds with 32 kHz stereo audio. Its decoded 120 video-frame hashes and all decoded audio-frame hashes are exactly identical to the accepted external four-node Pixel/VAE baseline. The prior colored line/plate and resize-ghosting artifacts are absent. Both baselines enter the same FL2VA Last Frame still interval at 3.666667 seconds; this is not an integrated-node regression.
|
||
- This acceptance is limited to the 1x5 FL2VA 2x case. Peak observed host use was approximately 47.7 GiB working set / 61.3 GiB private bytes with about 7.3 GiB system RAM still available; 3x5 Terminal Merge 2x remains a separate GPU/resource acceptance. V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, Advanced Second Pass, and the Latent Resize Utility were not changed by this Main backend implementation.
|
||
|
||
## Phase 2A P2-0 file-backed Tensor backend PoC (2026-08-24)
|
||
|
||
- P2-0 is accepted as a source-only Windows backend/lifetime PoC. It is not imported or registered by the Production node package, and neither ComfyUI_W nor ComfyUI_WAN runtime files were changed.
|
||
- The selected candidate is `torch.from_file(shared=True)`: it returns a contiguous CPU Tensor backed by a dedicated file while preserving the normal `IMAGE` Tensor contract.
|
||
- Backing files use strict allocation IDs plus JSON ownership sidecars. Cleanup is serialized per root across threads/processes, fails closed for invalid ownership metadata, retains live/cache-referenced storage, retries sharing violations, and removes dead-process stale files or eligible orphans.
|
||
- Storage lifetime is alias-aware through PyTorch `StorageWeakRef`. Future Production integration must use the process-global `get_process_lifetime_manager()`; a node-local manager is not an accepted integration contract.
|
||
- GC finalizers only enqueue cleanup intent. Filesystem I/O uses a consistent root-lock-before-manager-lock order, so cleanup/collect/finalizer interleavings cannot form the reproduced ABBA deadlock. Cancellation cleanup also protects a numeric file descriptor that has already been closed and reused by unrelated code.
|
||
- Primary validation: 26 focused tests and the standalone probe pass under ComfyUI_W Python/PyTorch; full repository validation is 334 passed. Secondary validation passes the same 26 tests and probe under ComfyUI_WAN Python 3.12.12 / PyTorch 2.13.0+cu130. Both probes are bit-exact, defer cleanup while an alias is alive, reclaim after the final alias is released, and leave zero managed backing files.
|
||
- This P2-0 acceptance does not claim end-to-end low-memory assembly. Real ComfyUI cache/requeue, Preview/VHS/Save consumption, RAM-vs-Disk assembler equivalence, and 9-15 GiB sequential-write/RSS behavior remain explicit later gates (P2-4/P2-5). Invalid sidecar quarantine/manual recovery is also not yet a public policy.
|
||
- P2-0 is now consumed by the independent V3.5 Production path described below. V3.4 remains fixed.
|
||
|
||
## Phase 2A P2-1 through P2-3 CPU acceptance (2026-08-24)
|
||
|
||
- P2-1 Memory Attribution is complete. V3.5 sequence reports capture chronological start, before CPU commit, after CPU commit, and completion snapshots; deduplicated retained AV/refine-context CPU/GPU bytes; projected natural/exact IMAGE bytes; process RSS/system availability; and read-only CUDA allocation/reservation counters. Collection is fail-soft and does not synchronize CUDA, reset peaks, clear caches, or alter V3.4 default behavior.
|
||
- P2-2 added `v3/file_backed_buffer.py`, backed by the accepted P2-0 `torch.from_file(shared=True)` lifetime contract. P2-6 now exposes `Auto` plus the original manual `RAM` and `Disk-backed` choices; there is still no silent Disk-to-RAM fallback. The first manager access performs root-scoped stale cleanup while live mappings remain protected by the ownership/lock contract.
|
||
- P2-3 adds the independent `v3/assembly_v35.py` and `v3/video_seam_v35.py` path used by `H3ContinuumAssembleSeamV35`. Video alone may be file-backed; Audio remains in RAM and preserves the V3.4 seam/sample-alignment result.
|
||
- Exact Duration allocates the final target-frame IMAGE shape first and writes trim, pad, and optional preserved final-frame spans directly. The assembler does not construct the final IMAGE with full-tensor `cat`, `stack`, `clone`, or `contiguous` copies. Audio remains intentionally RAM-backed and may use separate natural/target RAM buffers.
|
||
- Physical `decode_groups` retain their original order and terminal-prefix trim contract. Video Seam reuses the V3.4 analysis/formula/action semantics but copies only the affected 1-4 boundary frames plus required anchors; an analysis/patch error falls back transactionally to native frames.
|
||
- Disk-backed publication occurs only after assembly, reporting, and Driving Audio validation succeed. Unpublished early/late errors abort the backing pair. A published IMAGE remains valid while any Tensor/storage alias is retained, and the backing file becomes cleanup-eligible only after the final alias is released.
|
||
- P2-3 CPU acceptance covers RAM/Disk bit-exact output, Seam Off/Analyze/Auto/Auto 2, Exact Duration ON/OFF, trim/pad/preserve-final behavior, 1x5, 2x5 Terminal Merge, 3x5 Long Terminal Merge, 32 kHz cumulative audio rounding, Audio Seam, physical order/trim, file lifetime, stale recovery, and error cleanup. Its original Auto-rejection gate is superseded by the accepted P2-6 policy below.
|
||
- Canonical full validation is `412 passed` with only the known `pynvml` FutureWarning. `py_compile` and `git diff --check` pass. The ComfyUI_W runtime verifier passes registration, native PackedLayout, and Fixed 3x5 prompt-plan checks; its focused P2 regression is `141 passed`.
|
||
- Pre-change snapshots are `pre-phase2a-p2-1-memory-attribution-20260824_030858` and `pre-phase2a-p2-2-disk-backed-20260824_032923`; the latter contains 166/166 source files with zero SHA-256 mismatch at revision `d77a467796c048908f8b7b83736e226e736e0f98`, plus both runtime copies.
|
||
- Only the P2-1 through P2-3 target files were synchronized to primary ComfyUI_W; all 19 source-or-preserved-merge comparisons match SHA-256. Its pre-existing Video Reference tooltip addition was preserved. ComfyUI_WAN was deliberately not partially synchronized because its prior V3.5 dependency set is incomplete.
|
||
- P2-4 and the assembler-focused P2-5 acceptance are accepted below. Auto backend selection remains deliberately unimplemented; no new sampler/model generation acceptance is claimed by P2-5.
|
||
|
||
## Phase 2A P2-4 ComfyUI integration acceptance (2026-08-24)
|
||
|
||
- P2-4 is formally PASS using ComfyUI_W's installed `PromptExecutor` with the default CLASSIC `HierarchicalCache`, Core `SaveImage`/`PreviewImage`, and installed Video Helper Suite `VideoCombine`. The deterministic fixture is decoded CPU IMAGE/AUDIO only; no model, VAE, sampler, media input, CUDA allocation, or GPU generation is involved.
|
||
- `v3/assembly_v35.py` now polls ComfyUI's normal interrupt flag between bounded 8-frame direct-copy batches and immediately after the provisional output allocation. This changes only the independent V3.5 writer; V3.4 Assembly/Seam remains untouched. RAM and Disk-backed outputs remain bit-exact.
|
||
- A completed Disk-backed IMAGE remains alive after an intentional downstream error because the actual Core cache retains it. Requeue with unchanged assembler inputs reuses the same cached output and creates no additional backing pair. After cache eviction, the next allocation reclaims the released pair before creating its replacement, so managed pair count remains bounded; final managed files are zero.
|
||
- Core Save wrote 120 PNG frames and Core Preview wrote 120 temporary PNG frames from the mapped IMAGE. VHS streamed H.264 plus AAC from the same IMAGE/AUDIO. `ffprobe` verified 16x16, 120 frames, 24 fps, 5.000 seconds, and 32 kHz audio; this small geometry is intentional for compatibility/lifetime isolation.
|
||
- Core interrupt propagation is accepted: the probe raises the same `InterruptProcessingException` path used by ESC, PromptExecutor reports `execution_interrupted`, and the unpublished Disk-backed pair is removed. Separate-process crash/restart simulation leaves one stale pair and the next process removes it on first manager access.
|
||
- Reusable probe: `tools/p24_comfy_integration_probe.py`. Final runtime evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\P24-runtime-W-final-20260824_1105` (`p24_result.json`, exact prompt graphs, 120 Save frames, 120 Preview frames, and the VHS MP4).
|
||
- Validation: full canonical pytest `415 passed`; P2-focused ComfyUI_W runtime pytest `89 passed`; runtime registration/native PackedLayout/Fixed 3x5 prompt-plan verification, changed-file `py_compile`, and `git diff --check` pass. Only the known `pynvml` FutureWarning remains.
|
||
- Accepted pre-change snapshots are `pre-phase2a-p2-4-comfy-integration-source-verified-20260824_101909` (415 files, zero SHA-256 mismatch, revision `d77a467796c048908f8b7b83736e226e736e0f98`) and `pre-phase2a-p2-4-comfy-integration-comfyui-w-verified-20260824_101938` (324 files, zero mismatch). The standard snapshot script attempts failed only because the pre-existing `.pytest_cache` ACL denies traversal; those incomplete attempts are not accepted snapshots.
|
||
- Only `v3/assembly_v35.py`, the P2-4 probe, and its test file were synchronized to ComfyUI_W; source/runtime SHA-256 matches. P2-4 does not claim RAM-pressure/LRU cache modes, arbitrary downstream nodes, large-output memory behavior, manual browser UI interaction, or GPU/long-duration acceptance.
|
||
|
||
## V3.5 compact UI parity (2026-08-24)
|
||
|
||
- `web/project_id.js` now applies the established V3.4 compact/runtime-settings path to `H3ContinuumSamplerV35` and `H3ContinuumAssembleSeamV35`. No Python schema, socket, Sampling, Conditioning, Assembly, Seam, or Run Storage behavior changed.
|
||
- Sampler `Report Detail`, `debug`, and `show_preview` remain hidden widgets controlled by the existing H3-Continuum Settings entries. `strict_compatibility` remains loadable for legacy workflows but is hidden and forced `false` before queueing.
|
||
- `Auto Resume ID Override` remains always hidden and internally generated. `Run Name (Optional Override)` and `Regenerate From` are visible only when Run Storage is `Save + Auto Resume`; `Variation Nonce` is visible only for an explicit stored-chunk regeneration.
|
||
- The V3.5 assembler's `Report Detail` and `Exact Duration` controls use the same hidden Settings/fixed-contract path as V3.4.
|
||
- Accepted snapshots: source `pre-v35-compact-ui-source-verified-20260824_104010` (420 files, SHA-256 mismatch 0, excluding `.git` and inaccessible `.pytest_cache`) and ComfyUI_W `pre-v35-compact-ui-runtime-w-verified-20260824_104048` (330 files, mismatch 0, excluding `.pytest_cache`). The standard snapshot attempts are incomplete because that existing cache ACL denies traversal.
|
||
- Validation: dedicated frontend regression `2 passed`, JavaScript module syntax PASS, and full canonical pytest `417 passed`. The initial full-test attempts stopped only at inaccessible default pytest temp/cache paths; the canonical run used a fresh writable `D:\Codex\_test_temp` base. P2-5 is accepted below.
|
||
- Synchronized only `web/project_id.js` and its dedicated regression test to ComfyUI_W. Source/runtime SHA-256 matches (`EA79AF984EB7D615F0D79384C8B505442AF2B00CF7BABFD53E82DA2733927BCC` and `8B75C23DCCFD1AC302596187F2FBBD1C9B71FDDCB5E666AA4F9EEDFD8B5B4B93`).
|
||
|
||
## Phase 2A P2-5 long-output acceptance (2026-08-24)
|
||
|
||
- P2-5 is formally PASS for the manual `RAM` / `Disk-backed` V3.5 Assembly path. Testing was headless through ComfyUI_W's installed `PromptExecutor`; no browser, mouse, or front-end automation was used.
|
||
- The real 3x5 FL2VA Long Terminal Merge fixture reuses accepted Run Storage revision `1e30d8bf2154bbe7`: 3 logical chunks, 2 physical decode groups (`37T + 77T`), Core Video/Audio VAE Decode on GPU, 640x640, 360 frames, 24 fps, 15.000 seconds, and 32 kHz stereo audio. No new sampler/model generation was run.
|
||
- Independent fresh-process RAM and Disk-backed results are bit-exact. Final video SHA-256 is `95d5479bc5fee9c235ece7996c7c88dcc60981625dbedc330344403016ca9b9d`; audio SHA-256 is `eec0c6ff378c5bfcc3e71d02bab14230dea02a57e18ce1212f1dc1be4c88a3f9`. Exact Duration applies `362 -> 360` frames and `-2667` audio samples identically; Seam Auto and physical group order/trim match.
|
||
- Real-size Core Save and Preview each consumed all 360 mapped frames. VHS produced H.264/AAC at 640x640, 360 frames, 24 fps, 15.000 seconds, and 32 kHz. Disk-backed downstream peak RSS was about 12.10 GiB versus about 11.70 GiB for Assemble/hash alone; no catastrophic full-output duplicate was observed.
|
||
- Same-input requeue reports every node through assembler/digest cached, retains exactly one backing pair, and creates no additional pair. Cache release removes managed `.bin`/`.json` files. Core interrupt raised after backing allocation follows `execution_interrupted` and leaves zero managed files. A separate crash/restart recheck leaves one stale pair and removes it on first manager access in the next process.
|
||
- The 9.49 GiB stress uses a deterministic 1536x1536x360 float32 IMAGE while preserving the accepted two-group Long Terminal Merge plan. RAM and Disk-backed video/audio hashes are bit-exact; both are finite; Disk-backed holds one 10,192,158,720-byte file while cached and leaves zero managed files after release.
|
||
- On Windows, RSS/working set is not the acceptance metric for file-backed savings because resident mmap pages are counted in RSS. At the stress digest boundary, RAM and Disk-backed RSS were both about 10.44 GiB. The correct commitment metrics show the benefit: RAM private/USS were 13,322,362,880 / 11,045,199,872 bytes, while Disk-backed private/USS were 3,122,954,240 / 855,834,624 bytes. File backing therefore removed about 9.50 GiB of anonymous private commitment and about 9.49 GiB of unique resident commitment for a 9.49 GiB output.
|
||
- Reusable probe: `tools/p25_gpu_acceptance_probe.py`; focused P2-5/P2-4/P2-3 regression is `58 passed`, and full canonical pytest is `420 passed` with only the known `pynvml` FutureWarning. `py_compile` and `git diff --check` pass.
|
||
- Accepted snapshots are `pre-phase2a-p2-5-gpu-acceptance-source-verified-20260824_105605` (422 files, mismatch 0) and `pre-phase2a-p2-5-gpu-acceptance-runtime-w-verified-20260824_105626` (331 files, mismatch 0). P2-5 changed only the source-side probe, its tests, and project records; no Production runtime sync was required.
|
||
- Primary evidence roots: `P25-long-terminal-ab-20260824_110031`, `P25-downstream-disk-20260824_110515`, `P25-interrupt-disk-20260824_110815`, `P25-cache-requeue-disk-20260824_111126`, `P25-stress-1536-disk-private-20260824_111836`, `P25-stress-1536-ram-private-20260824_111907`, and `P25-lifetime-recheck-20260824_111947` under `D:\Codex\_test_results\ComfyUI-H3-Continuum`.
|
||
- Auto backend is accepted in P2-6 below. Arbitrary downstream nodes may still materialize their own full copies. New 3x5 sampler/model generation and 3x5 Main Hi-res Fix 2x GPU/resource acceptance remain separate work; P2-5 validates the assembler using already accepted physical latents and real Core GPU Decode.
|
||
|
||
## Phase 2A P2-6 Auto Buffer Backend acceptance (2026-08-24)
|
||
|
||
- P2-6 is formally PASS. `H3ContinuumAssembleSeamV35` now offers `Auto`, `RAM`, and `Disk-backed`; `Auto` is the default for newly created V3.5 assembler nodes. Existing saved string values for either manual backend remain valid and retain their exact manual behavior.
|
||
- Auto selects RAM only when the final IMAGE is at most 4 GiB and available physical memory is at least `final IMAGE + max(4 GiB, 10% of total physical RAM)`. Larger outputs or insufficient headroom select Disk-backed. If memory counters are unavailable, Auto chooses Disk-backed rather than assuming RAM is safe.
|
||
- A Disk-backed Auto decision requires free space for the final IMAGE plus a fixed 2 GiB disk reserve. Capacity-probe failure, insufficient disk, or allocation failure is explicit and never falls back to RAM. Users may still select manual RAM when they independently know it is safe.
|
||
- The V3.5 report records requested/resolved backend, final IMAGE bytes, available/total RAM, computed RAM reserve, 4 GiB RAM limit, disk free/reserve values, and the decision reason. Manual paths do not query or change Auto policy.
|
||
- Real headless acceptance `P26-auto-ram-long-terminal-20260824_113400` reused the accepted 3x5 FL2VA Long Terminal Merge revision and Core GPU Decode. The 1,769,472,000-byte output selected `Auto -> RAM` with about 39.2 GiB available and a 6.39 GiB RAM reserve. Its 640x640 / 360f / 15.000s video and 32 kHz audio hashes exactly match P2-5.
|
||
- Large-output acceptance `P26-auto-disk-stress-20260824_113610` used the 10,192,158,720-byte / 9.49 GiB IMAGE fixture and selected `Auto -> Disk-backed` because the final IMAGE exceeded 4 GiB. It retained one backing pair while cached, returned the accepted stress video/audio hashes, and left zero managed files after release.
|
||
- Focused Auto/P2 regression is `66 passed`; full canonical pytest is `428 passed`; changed-file and runtime `py_compile`, runtime registration/native PackedLayout/Fixed 3x5 verification, and `git diff --check` pass. Only the known `pynvml` FutureWarning remains.
|
||
- Accepted snapshots are `pre-phase2a-p2-6-auto-source-verified-20260824_112744` (426 files, mismatch 0) and `pre-phase2a-p2-6-auto-runtime-w-verified-20260824_112744` (331 files, mismatch 0), excluding the pre-existing inaccessible `.pytest_cache`. Only `v3/file_backed_buffer.py`, `v3/assembly_v35.py`, and the Auto portion of `v3/driving_nodes.py` were synchronized to ComfyUI_W; its pre-existing Video Reference tooltip addition remains preserved.
|
||
- V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, and every V3.4 socket remain unchanged. Auto does not guarantee that arbitrary downstream nodes avoid their own full copies. New 3x5 sampler/model generation and 3x5 Main Hi-res Fix 2x resource acceptance remain separate.
|
||
|
||
## V3.5 release-candidate preparation (2026-08-24)
|
||
|
||
- Public package version, Python package metadata, `metadata.ini`, Changelog, English/Japanese README, package information, and validation text now identify `3.5.0` consistently.
|
||
- Public scope is explicit: Sampler V3.5 refine-context, Advanced Second Pass Bridge, and Assemble + Seam V3.5 are release paths; integrated Main Hi-Res Fix remains Experimental. The standalone Latent Resize node is documented as an advanced utility rather than the recommended Main 2x backend.
|
||
- Added `docs/V35_HIRES_FIX.md` with standard Main and external-upscaler wiring, separate First/Second Pass scheduler roles, the accepted `simple / 10 steps / denoise 0.35` baseline, OFF passthrough, audio behavior, and current GPU limits. No diagnostic UI workflow is published as an accepted V3.5 template.
|
||
- Distribution Manifest includes the V3.5 Production modules, CPU regressions, release metadata test, public guide, and the Production-imported P2-0 backend. P2-4/P2-5 runtime probes remain repository validation tools rather than distribution Manifest payloads. No credentials or generated test media are included.
|
||
- Release metadata regression and targeted V3.5 tests pass (`75 passed`). Full canonical pytest is `430 passed`; only the known `pynvml` FutureWarning remains. Source and ComfyUI_W runtime verifier both report package `3.5.0`, native PackedLayout PASS, Fixed 3x5 prompt-plan PASS, and the complete expected node registry.
|
||
- Representative headless GPU revalidation `V35-release-candidate-auto-long-terminal-20260824` is PASS. Accepted 3x5 FL2VA Long Terminal Merge latents decoded through Core GPU Video/Audio VAE and Assemble V3.5 Auto as 3 logical / 2 physical groups (`37T + 77T`), 640x640, 360 frames, 24 fps, 15.000 seconds. Auto selected RAM for 1,769,472,000 bytes; video/audio SHA-256 exactly match P2-5/P2-6, all outputs are finite, and managed backing files after release are zero.
|
||
- Remaining publication limits: Main Hi-Res Fix 3x5 2x and Reference/Hybrid-specific Second Pass GPU acceptance are pending; arbitrary downstream full IMAGE copies are outside Continuum's guarantee. Existing V3.4 workflows remain supported.
|
||
- Verified release-preparation snapshots are `pre-v35-release-preparation-source-verified-20260824_115858` (426 files, mismatch 0) and `pre-v35-release-preparation-runtime-w-verified-20260824_115858` (331 files, mismatch 0). Commit/push remain pending explicit user instruction.
|
||
|
||
## V3.5 pending GPU acceptance closure (2026-08-24)
|
||
|
||
- Real ComfyUI API acceptance used an NVIDIA GeForce RTX 5060 Ti with 16,311 MiB VRAM. Production code was not changed.
|
||
- Main Hi-Res Fix `3 x 5s` FL2VA, 576x576 to 1152x1152 Lanczos, `res_multistep`, workflow `simple / 10 steps / denoise 0.35`, preserved 3 logical chunks and 2 physical groups (`37T + 77T`). First Pass and the 37T group's 10-step Second Pass completed. The terminal logical `[2,3]` / 77T group failed at its first Second Pass inference with CUDA OOM: 15.93 GiB device limit, 10.04 GiB allocated, 1.49 GiB requested, zero CUDA free. This condition is formally not accepted on the tested 16 GiB GPU; it is a resource boundary, not evidence that physical grouping changed.
|
||
- Hybrid FL2VA + Reference `1 x 5s`, 576x576 to 1152x1152, passed the integrated Main Hi-Res Fix. The node used `conditioning_source=refine_context`, one physical group / one sampling pass, actual refine seed `2946484852083492747`, output latent `(1, 24, 37, 72, 72)`, temporary refined audio discard, and bit-exact first-pass audio LATENT passthrough. Core Decode / Assemble produced 1152x1152, 120 frames, 24 fps, 5.000 seconds with 32 kHz stereo audio; visual inspection found no persistent colored line/plate artifact, flash, or temporal collapse. Output SHA-256 is `E5B27CABCD5FBC1422F169AC6E75570EF4D54010CD04A3439852D8DED15C1384`.
|
||
- The same Hybrid/Reference input also passed the public `H3ContinuumSecondPassV35` node directly at 576x576: `conditioning_source=refine_context`, actual refine seed `5765580629374593532`, shape `(1, 24, 37, 36, 36)`, one sampling pass, bit-exact first-pass audio passthrough, Core Decode / Assemble 120 frames / 5.000 seconds, and clean contact-sheet inspection. Output SHA-256 is `67311A797CD812584B4E2CCC21D7DF83AE8C1F1D72AB8BF5FA1C142D00894D2D`.
|
||
- The Hybrid prompt used the plain text `Picture 3` instead of the recommended explicit `<Picture 3>` tag, so preflight emitted a non-blocking H3C-P103 warning. The connected reference still participated in First Pass conditioning; the run proves transport/execution acceptance, not a warning-free public prompt example.
|
||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\V35-pending-gpu-acceptance-20260824`. Remaining limits are Main 3x5 2x on larger VRAM or a lower-memory terminal-group implementation, and Reference/Hybrid durations longer than 1x5.
|
||
- Pre-documentation snapshots are `pre-v35-pending-gpu-results-source-verified-20260824_130824` (431 files) and `pre-v35-pending-gpu-results-runtime-w-verified-20260824_130824` (334 files), excluding `.git` and the pre-existing inaccessible `.pytest_cache`.
|
||
|
||
## V3.5.0 formal release documentation (2026-08-24)
|
||
|
||
- V3.5.0 is the current formal release. The public README now leads with the two major additions: Continuum-aware Second Pass / experimental integrated Hi-Res Fix, and low-memory Assemble + Seam V3.5.
|
||
- V3.4 node IDs remain registered intentionally for saved-workflow compatibility. V3.5 does not replace V3.4 Sampling, Conditioning, Terminal Merge, Assembly, Seam, Run Storage, or public sockets; existing V3.4 workflows can continue unchanged.
|
||
- Added separate public node images `docs/images/v35-hires-fix-node.png` and `docs/images/v35-second-pass-node.png`. BasicScheduler is omitted visually but the README explicitly requires workflow-side SIGMAS and records the accepted `res_multistep`, `simple / 10 steps / denoise 0.35` baseline. The Main image shows the accepted Lanczos 2x selection.
|
||
- Public memory claims are limited to measured Windows host-memory behavior: for the 9.49 GiB 1536x1536x360 stress IMAGE, Disk-backed reduced private memory by about 9.50 GiB and USS by about 9.49 GiB while preserving exact video/audio hashes. No sampler GPU-VRAM reduction is claimed.
|
||
- Public GPU limits now distinguish accepted 1x5 576-to-1152 Main and Hybrid/Reference paths from the unaccepted 3x5 2x RTX 5060 Ti 16 GiB condition, which failed at the terminal 77T group's first Second Pass inference after the 37T group completed.
|
||
- Final source validation is `430 passed`; all 159 Manifest entries verify by SHA-256; `git diff --check` passes. ComfyUI_W runtime verifier passes package 3.5.0 registration, native PackedLayout, and Fixed 3x5 planning. Nine changed public documentation/assets match the snapshotted runtime by SHA-256.
|
||
- Verified pre-release snapshots are `pre-v35-formal-readme-release-source-verified-20260824_141358` (431 files) and `pre-v35-formal-readme-release-runtime-w-verified-20260824_141359` (334 files), excluding `.git` and the pre-existing inaccessible `.pytest_cache`.
|
||
|
||
## Issue #11 documentation clarification (2026-08-25)
|
||
|
||
- `Reference Size` and `Video Guide Size` are existing V3.5 controls, not missing inputs. Compact UI reveals them only after a corresponding Reference Image or Video Guide Frames connection.
|
||
- `Video Guide Size` is the V3.5.1 display name for the earlier `Video Reference Size`; the backend key and saved-workflow compatibility are unchanged.
|
||
- Reference Images and Video Guide Frames are resized internally before VAE encoding. External resize remains optional for deliberate crop, forced source upscaling, or a custom algorithm; loader-side decoding and frame-rate conversion remain unchanged.
|
||
|
||
## V3.5.1 public-state cleanup (2026-08-25)
|
||
|
||
- Project instructions now identify V3.5.1 as the current public baseline while retaining V3.4 node IDs, public contracts, and saved-workflow compatibility.
|
||
- The later real-browser acceptance superseded the earlier simulated Last Queued Seed result; that experiment is withdrawn and standard ComfyUI Seed behavior is restored.
|
||
- GitHub Issues #9, #10, and #11 received implementation guidance and are closed as completed. Issue #11 was closed after the conditional-UI/internal-resize README clarification reached public `main`.
|
||
|
||
## GitHub Actions dependency/stub repair (2026-08-25)
|
||
|
||
- The public V3.5.1 Actions failure was limited to the isolated CI environment: three P25 tests lacked `psutil`, and two Hi-Res Fix tests lacked ComfyUI Core's `comfy_extras.nodes_minimax_h3` module.
|
||
- `psutil` is now an Actions-only test dependency. Runtime `requirements.txt` remains dependency-free.
|
||
- `tests/test_v35_hires_fix.py` provides a file-local `comfy_extras.nodes_minimax_h3` stub containing only Core's `CANVAS_MULTIPLE = 32`. Production Core imports and Hi-Res Fix behavior are unchanged.
|
||
- Focused Hi-Res Fix/P25 validation is `16 passed`; full validation is `450 passed` with only the known `pynvml` FutureWarning. The first full local run encountered only the existing Windows global pytest temp ACL error; the accepted rerun used a dedicated writable `--basetemp`.
|
||
|
||
## Issue #13 contrast / oily drift investigation (2026-08-27)
|
||
|
||
- Controlled 5x5-second I2VA GPU comparison used one fixed First Image, base seed `130013`, 384x384, Balanced 22, Audio Continuity ON, `res_multistep` / `simple` 8 steps, Sage Attention, no LoRA, Spectrum OFF, Run Storage OFF, and both seams OFF.
|
||
- Fixed, same-text List, and same-text Timeline resolved to the same five prompt strings and SHA-256 `9d95f563bf519ff548ac61f65d4f232d584f3531fdc4c3d2468199921b806326`. Their decoded RGB video MD5 (`1e655f35cfb7b93082831d275e3ee6b5`) and PCM audio MD5 (`0dc94fcd389e2e0da5342a88d2978327`) are identical. Fixed mode is not the cause.
|
||
- V3.6 Standard Masked AV did not reproduce the reported cumulative contrast/saturation pattern. Chunk 5 minus Chunk 1 was luma `+0.0688`, saturation mean `-0.0493`, saturation p95 `-0.4833`, and contrast `+0.0039`.
|
||
- Compatibility / Reference Context did reproduce the relevant pattern under the matched condition: luma std `+0.0472` (about 25.9%), saturation mean `+0.0207` (about 16.7%), saturation p95 `+0.0773`, and contrast `+0.0940` (about 14.9%) from Chunk 1 to 5, while mean luma stayed nearly stable.
|
||
- A varied List requesting restrained contrast/color changed the output but did not improve the measured drift over the same-text Standard case. The evidence therefore points to legacy Reference Context visual self-conditioning rather than Prompt Format or repeated identical text itself.
|
||
- Current workaround: use V3.6 Standard + Balanced 22 for the accepted Masked AV transport; avoid Compatibility for long runs showing oily/contrast drift. V3.6.1 non-Balanced Audio Continuity fallback still intentionally uses Reference Context, so the same model-side risk can remain for Fast 5 / Strong 39 / Auto.
|
||
- No Production change is justified in V3.6.1. Further Reference Context normalization or dynamic 5f/39f Masked AV belongs to V3.6.2-or-later research with separate quality gates.
|
||
- Evidence root: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-drift-20260827_204827`; report `ISSUE_13_INVESTIGATION_REPORT.md`, five-case contact sheet, full CSV/JSON metrics, prompts, histories, and videos are retained. Diagnostic tools are isolated under `tools/` and not imported by Production.
|
||
- Validation: diagnostic `py_compile` PASS; full pytest `543 passed` with the known `pynvml` FutureWarning; ComfyUI_W V3.6.1 runtime verifier PASS; `git diff --check` PASS. No commit/push/release was performed.
|
||
|
||
## Audio Continuity Research Gate (2026-08-27/28)
|
||
|
||
- Audited the existing 24 fps / 40 Hz timeline contract without changing Production. Regression coverage now fixes 124f/207T (+1/3 tick), 141f/235T (0), 260f/433T (-1/3 tick), 2/3/5-chunk cumulative PCM alignment, State 39f/65T capacity, Audio40 end alignment, and Seam diagnostic non-interference.
|
||
- A private tools/diagnostic-only A/B compared the accepted Balanced 22 Reference Context (Audio 37T) with an independently end-aligned 40T audio history. Three fixed-seed 3x5-second T2VA pairs covered music, dialogue, and ambient sound at 576x576 / 20 steps / Seam OFF. All six completed with finite latents, 360f / 15.000s / 32 kHz stereo output, empty final queue, and no visible video regression.
|
||
- 40T adds six packed rows (+0.036%) and showed no material VRAM regression. Continuation Sampling deltas were -5.41%, -0.51%, and -1.82%, but no speed claim is accepted because the larger layout cannot explain that direction and execution-order/system variance remains.
|
||
- Final audio-boundary metrics were content-dependent: Ambient improved, Dialogue mixed immediate-jump improvement with RMS/floor regressions, and Music worsened both immediate boundary jumps (including 0.33319 -> 0.44815 at 243f). Therefore 40T did not meet the cross-content promotion criterion.
|
||
- Final decisions: **KEEP CURRENT 37 TICKS**; adopt Seam diagnostics only as research tooling and keep them out of Production UI; adopt timeline-hardening tests without a new Production abstraction. Evidence and the detailed report are under `D:\Codex\_test_results\ComfyUI-H3-Continuum\audio-continuity-research-20260827_2335`.
|
||
- Validation: focused `33 passed`, related contract suite `115 passed`, full pytest `552 passed`, compileall/runtime verifier/diagnostic source-runtime SHA/`git diff --check` PASS. Production, frontend, schema, version, commit, push, and release were unchanged. ComfyUI_W remains running.
|
||
|
||
## V3.8 FL2VA Terminal Smart Regenerate Phase F-G4 STOP (2026-08-31)
|
||
|
||
- G4-A Full established the intended FL2VA 3x5-second baseline: 3 logical chunks, 2 physical groups, and one Terminal Merge physical sample for Chunks 2–3. Output was 544x544, 360 frames, 24 fps, 15.000 seconds, and 32 kHz stereo; peak VRAM was 15.375 GiB.
|
||
- G4-B Review Queue 1 correctly generated only Chunk 1, saved `review_ready` unit `1-1`, performed one physical sample, retained the Last Frame contract without applying it, and produced a valid 120-frame / 5-second partial output. The sampler executed and was not cached.
|
||
- Mandatory Full/Review Chunk 1 parity failed. Seed, prompt hash, plan, revision, and full sampling-contract SHA were exact, but video latent SHA differed (`d6da9472...` vs `140064e6...`) and audio latent SHA differed (`dab3e1ee...` vs `4742f4d9...`). This is a real tensor mismatch, not a safetensors container-header difference.
|
||
- The failure boundary is the per-execution FL2VA image-conditioning path: First Image Video VAE encoding and Qwen3-VL image conditioning are recomputed for separate queues. Existing evidence does not hash those two intermediate outputs, so the exact subpath is not yet proven. Terminal grouping, Review cap, seed/prompt identity, Last Frame timing, Run Storage identity, and sampler cache are excluded by the captured evidence.
|
||
- Stopped before Queue 2, Terminal Smart Regenerate, nonce, restart, G5, or post-PASS regression. Production code was not changed. Next approval should be for a focused encoder reproducibility diagnostic, followed only then by the smallest branch-specific cache/determinism fix.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g4-20260831_143138\PHASE_F_G4_STOP_REPORT.md`. Snapshot `pre-v38-review-phase-f-g4-terminal-smart-xhigh-gpu-gate-20260831_142609-0a7c670` contains 724 exact files excluding `.git`/`.pytest_cache`. Dedicated 8190 was stopped after Queue empty; 8188 was not operated.
|
||
|
||
## V3.8 G4 FL2VA Conditioning Determinism diagnostic (2026-08-31)
|
||
|
||
- The observation-only, no-Sampling G4 diagnostic classified the image-conditioning path as **CASE E**: P0 source image, P1 Continuum-resized image, P2 First Image Video VAE latent, P3 Qwen image input, P4 token payload, P5 CLIP conditioning, P6 metadata, and P7 keyframe-attached final conditioning were all bit-exact.
|
||
- Exact parity held for cold vs warm repeat, Full-preparation vs Review-preparation, and cold vs backend restart. P0 was `1x1344x1024x3` float32 CPU; P1/P3 `1x544x544x3`; P2 `1x24x1x34x34`; the P5 tensor was `1x330x5120` float32 CPU. Structured P4-P7 capture contained no unsupported/opaque types.
|
||
- Cold/restart CPU-to-CUDA loading and restart-created patch UUIDs did not affect outputs. VAE has no object patches/wrappers; CLIP retained only its existing `manual_cast_dtype` patch, no wrappers, and stable empty tokenizer options. Deterministic algorithms and cuDNN deterministic mode remained OFF, yet P0-P7 were exact.
|
||
- VAE/CLIP caches are not justified by this evidence. The remaining first observed difference is the post-Sampling Chunk 1 latent from G4. Any next diagnostic must move to target/noise/mask/sigmas, first H3 model-call input/output, and final sampler latent under the original Full Terminal then Review ordering; Production should remain unchanged until that first divergence is proven.
|
||
- Added only evidence tooling `tools/v38_g4_conditioning_determinism_gate.py`; the collector was loaded from an external extra custom-node path. Production Critical files remain source/runtime/snapshot-identical. Dedicated 8190 was restarted once and stopped after Queue empty; 8188 was not operated.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-g4-conditioning-determinism-20260831_150837\G4_CONDITIONING_DETERMINISM_STOP_REPORT.md` and `g4_conditioning_determinism_summary.json`. Snapshot `pre-v38-g4-conditioning-determinism-xhigh-gate-excluding-pytest-cache-20260831_150837-0a7c670` contains 724 exact files.
|
||
|
||
## V3.8 G4 Sampling Determinism diagnostic (2026-08-31)
|
||
|
||
- The Extra High Sampling-boundary Gate completed the A-D order matrix without changing Production: Fresh Full vs Fresh Review, Full -> Review on one backend, and Review -> separate Review on one backend. S0 input AV latent, S1 fixed latent, S2 noise, S3 mask, S4 SIGMAS, S5 seed, S6 sampler signature, S7 final conditioning, M0 first model x, M1 timestep, M2 context payload, M3 first prediction, and the final sampled Video/Audio latent were exact in every tensor comparison. S7 also matched the prior P7 hash for all six observed runs.
|
||
- C and D changed only S8 warm-runtime residency: the first run entered Sampling with the model on CPU and no current weight-patch UUID, while the next run entered with the model on `cuda:0` and a current weight-patch UUID. M0-M3 and final AV remained exact, so this is an observed warm state transition, not a reproduced bit-divergence boundary.
|
||
- A current all-custom-nodes, public `H3ContinuumSamplerV38`, no-hook control reproduced the historical Full hashes (`d6da9472...` video / `dab3e1ee...` audio), but its following Review Queue 1 matched Full exactly instead of reproducing the historical `140064e6...` / `4742f4d9...` mismatch. The old G4 divergence is therefore currently non-reproducible; root cause remains unresolved and no Production fix is justified. The old runner automatically started Queue 2 after Queue 1 passed; that out-of-scope dedicated-8190 queue was interrupted immediately and is not treated as a product failure.
|
||
- A first version of the collector hashed CUDA M0-M3 immediately after the first forward and was rejected because the host synchronization changed execution timing. Accepted evidence retains tensor references and hashes them only after Sampling. Current no-hook Full/Review hashes match the accepted collector Full/Review hashes, and every runtime hook restored successfully.
|
||
- Step tracing and accelerator ON/OFF were not run because no sampled-tensor or model-forward divergence was reproduced. The minimal next action, only if approved, is a repeated exact C reproduction matrix that preserves the public Full/Review cache contract; do not add caches, force deterministic modes, reload models, or disable accelerators without a reproduced first divergence.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-g4-sampling-determinism-20260831_154505\G4_SAMPLING_DETERMINISM_STOP_REPORT.md`, `g4_sampling_determinism_summary.json`, and `no_hook_control\gpu_r1\g4_summary.json`. Snapshot `pre-v38-g4-sampling-determinism-xhigh-gate-excluding-pytest-cache-20260831_154505-0a7c670` contains 725 exact files excluding only `.git`/`.pytest_cache`.
|
||
- Added only `tools/v38_g4_sampling_determinism_gate.py` and an evidence collector under `tools/g4_sampling_diagnostic_node/`; the loaded collector copy remained outside the Production runtime. Validation passed `835` tests, Manifest `217/217`, runtime verifier, Python/JavaScript syntax, and `git diff --check`. Dedicated 8190 and its launcher were stopped after Queue empty. Port 8188 remained HTTP 200 during the diagnostic, but a final read-only post-report check found no listener; this task never stopped, restarted, or otherwise operated 8188. No commit, push, version, tag, or release was performed.
|
||
|
||
## V3.8 G4R FL2VA Terminal Repeatability final PASS (2026-08-31)
|
||
|
||
- The hook-free public Production path passed three independent Fresh-backend Full -> new Review Queue 1 cycles exactly. R1/R2/R3 all matched seed `16802892808542812700`, Prompt hash `c7ec4b70...`, Video latent `d6da9472...`, and Audio latent `dab3e1ee...`; each Review generated only Chunk 1, retained `review_ready` unit `1-1`, and did not apply Last Frame. No additional repeat cycles were run.
|
||
- One separate Fresh final cycle passed Full, Review Queue 1, and Review Queue 2. Queue 2 reused Chunk 1 without Sampling, generated Chunks 2-3 as one Terminal physical group, committed both logical entries to one revision, and completed. All three logical seed/Prompt/Video/Audio/plan records, assembly plan, physical decode groups, decoded RGB, decoded 32 kHz stereo PCM, 360 frames, 24 fps, and 15.000-second output matched Full exactly.
|
||
- One Terminal `Regenerate Current` passed with target/boundary Chunk 2, Chunk 1 exact reuse, nonce `0 -> 1`, physical seed `8293933729523758687 -> 8951320726233541764`, both logical variations changed, both entries owned by new revision `baf61e416796683f`, and no old/new half mix. The evidence runner initially stopped only because it incorrectly required the `terminal merged pair is atomic` reset sentence. That sentence is emitted for a partially preserved Terminal pair; direct Smart boundary 2 has no partial pair, so the assertion was corrected outside Production and existing GPU evidence was reassessed without rerunning.
|
||
- Final classification is **Phase F-G4 = PASS with prior transient non-reproducible parity event**. The original mismatching Review hashes remain retained as a known historical event; no Production fix, deterministic forcing, model reload, attention/cache/dtype change, or diagnostic hook was applied.
|
||
- Peak VRAM was `15.54 GiB`, peak process RSS `47.50 GiB`, and minimum available RAM `4.38 GiB`. Validation passed `835` tests, Manifest `217/217`, runtime verifier, Python `173/173`, JavaScript `7/7`, eight source/runtime SHA pairs, and `git diff --check`. Snapshot `pre-v38-g4r-repeatability-xhigh-gate-excluding-pytest-cache-20260831_175742-0a7c670` contains 731 exact files excluding only `.git`/`.pytest_cache`. Dedicated 8190 was stopped after Queue empty; 8188 remained listening and was not operated. G5, commit, push, version, tag, and release were not performed.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g4r-20260831_175742\PHASE_F_G4R_PASS_REPORT.md` and `g4r_final_summary.json`.
|
||
|
||
## V3.8 Phase F-G5 Driving Audio Review / Smart Regenerate PASS (2026-08-31)
|
||
|
||
- The public, hook-free `H3ContinuumSamplerV38` path passed the required 20-second T2VA Driving Audio Gate at `2 × 10 s`, fixed Prompt/seed/Sampler/SIGMAS/resolution, and Run Storage `Save + Auto Resume`. The fixed input is `H3_Easy_Driving_Audio_Gate_21s.flac`, SHA-256 `ad9d942c...`, FLAC 32 kHz stereo, 20.974938 seconds. Reference Audio, Reference Image, Video Guide, and Still Guide were disconnected.
|
||
- Full used two physical samples. Review Queue 1 generated only Chunk 1 and saved `review_ready` unit `1-1`; Queue 2 exactly reused Chunk 1 and sampled only Chunk 2. Both logical seed/Prompt/Video latent/Audio latent/plan records, assembly, decoded RGB, and final Driving Audio PCM matched Full exactly. Output was 480 frames, 24 fps, 20.000 seconds, 32 kHz stereo, and 640,000 samples/channel. Final PCM SHA was `ea81cc52...`; the 10-second boundary was finite with no zero-RMS 200 ms window.
|
||
- `Regenerate Current` targeted Chunk 2, exactly reused Chunk 1, advanced nonce `0 -> 1`, and produced a new Chunk 2 Video/Audio latent variation. The final PCM remained exactly `ea81cc52...`, confirming the existing Driving Audio replacement contract. After a fresh 8190 restart, canonical revision `dfdadc00aa25c4a7`, boundary 2, nonce 1, unit `2-2`, and complete head were restored.
|
||
- A separate post-restart Review Queue 1 -> `Finish Remaining` run passed: Chunk 1 was exactly reused, only Chunk 2 was sampled, and Full latent/assembly/decoded video/final PCM parity remained exact. All Review actions executed the Sampler without ComfyUI cache; no OOM, NaN/Inf, Decode, Assembly, dropout, or duration error occurred.
|
||
- The optional 30-second branch-Continue auxiliary Gate was not run because the fixed accepted source is 20.974938 seconds; no replacement test audio was introduced. Peak VRAM was `15.56 GiB`, peak RSS `45.85 GiB`, and minimum available RAM `5.01 GiB`; the 16 GiB Gate passes with small headroom.
|
||
- Regression passed `835` tests, Manifest `217/217`, runtime verifier, Python `173/173`, JavaScript `7/7`, source/runtime SHA `8/8`, and `git diff --check`. Snapshot `pre-v38-review-phase-f-g5-driving-audio-gpu-gate-excluding-pytest-cache-20260831_203817-0a7c670` contains 738 exact files excluding only `.git`/`.pytest_cache`. Production code was unchanged. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-review-phase-f-g5-20260831_203817\PHASE_F_G5_PASS_REPORT.md` and `g5_summary.json`.
|
||
- Final classification: **Phase F-G5 Review / Smart Regenerate × Driving Audio GPU Gate = PASS**. Final Integration Gate, commit, push, version, tag, and release were not performed.
|
||
|
||
## V3.8 Reliability Phase 1-4 GPU baseline (2026-09-01)
|
||
|
||
- Restarted only the ComfyUI_W backend on `127.0.0.6:8188`; the Stability Matrix frontend was not stopped or changed. The accepted no-guide and Video Guide smoke conditions were V3.8, 2x5 seconds, 576x576, fixed seed, Balanced 22 continuity, Audio Continuity ON, Run Storage Off, and 8 Sampling steps.
|
||
- Both runs completed as 240 frames / 10.000 seconds / 24 fps / 32 kHz stereo with two physical groups and no reliability advisory. The Video Guide run reported 124 -> 124 alignment, PASS video/audio grids for both groups, native Core contracts, and relative visual load `1.235` / Medium for the resolved 640x640 124-frame guide.
|
||
- Peak observed device use was 15.42 GiB without Video Guide and 15.25 GiB with Video Guide. Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-reliability-phase1-4-gpu-baseline-20260901`; the accepted pre-Phase5 source snapshot is `v38-phase1-4-gpu-baseline-pass-pre-phase5-excluding-pytest-cache-20260901_015207-0a7c670` with 747 exact files.
|
||
|
||
## V3.8 Phase 5 Reference Encode Cache (2026-09-01)
|
||
|
||
- Added a V3.8-only, process-local, Video-VAE-scoped reference encode LRU. Reference Images and Video Guide use existing content identities; safe CPU Reference Audio is supported. Private CPU masters are detached/contiguous and cloned on every hit. GPU/meta results bypass storage. Limits are 16 entries / 512 MiB, with no disk, Session, State, Run Storage, workflow, or public-schema identity.
|
||
- Legacy V3.7 and older paths remain cache-disabled. Full Run Storage reuse still bypasses all reference encode/cache work. First/Last Frame, Still Guide, Driving Audio, Timeline Video, Prompt/CLIP, Sampling, and decode are not cached by this feature. Detailed reports expose only aggregate cache actions and resident limits.
|
||
- CPU/static validation passed 898 tests after adding a regression for the V3.8-to-Production internal flag chain; focused validation passed 56. Manifest is 223/223. The default pytest temp root had inherited ACL denial, so accepted full validation used a fresh `D:\Codex\_test_temp` base.
|
||
- The first GPU submission stopped before Sampling because the Production facade omitted the new internal keyword from one method signature. The minimal internal-only forwarding fix was applied; no public schema or generation contract changed.
|
||
- Fresh-process GPU A/B then passed with 2x5 seconds, 576x576, fixed seed, Balanced 22, Audio Continuity ON, Video Guide Efficient 0.4 MP, Run Storage Off, and 8 steps. Cold A was `miss=1/store=1` with 26.656295 seconds of Video Guide conditioning; Warm B was `hit=1` with 0.006424 seconds. Both retained one 5.4 MiB CPU cache entry.
|
||
- Cold/Warm matched exactly for both raw Video latent tensor SHA-256 values, both raw Audio latent tensor SHA-256 values, logical seeds, normalized Assembly Plan, two physical Sampling groups, and the 576x576 / 240-frame / 10-second / 24-fps / 32-kHz-stereo media contract. API time was 472.253 / 437.818 seconds; peak observed device use was 15.11 GiB. No OOM, NaN/Inf, advisory, or Prompt warning occurred.
|
||
- Evidence is under `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-ref-encode-cache-gpu-gate-20260901`. Accepted Phase5 source snapshot: `v38-phase5-ref-encode-cache-pass-excluding-pytest-cache-20260901_024700-0a7c670`. Phase 6 must remain read-only and may begin only from this accepted baseline.
|
||
|
||
## V3.8 Phase 6 Chunk Boundary Diagnostic (2026-09-01)
|
||
|
||
- Extended only `v3/reliability_v38.py` with a post-generation structural boundary diagnostic and added `tests/test_v38_chunk_boundary_diagnostics.py`. It reads Assembly Plan metadata plus physical output counts; it does not read Tensor values, clone/copy/transfer tensors, decode media, hash tensors, invoke ContextDiagnosticsTracker, or change Sampling/Conditioning/Run Storage/Assembly.
|
||
- Logical visible ranges validate `total_frames - trim_frames == net_frames`, cumulative and recorded frame ranges, native `context_slots()` / `audio_latent_t()` values, and logical-to-physical mapping. Basic remains silent on PASS. Detailed emits one structural-only header and one line per boundary.
|
||
- Focused validation passed 53 tests. It covers zero/one/two boundaries, Fast 5 (`2T/8T`), Balanced 22 (`7T/37T`), Strong 39 (`12T/65T`), invalid geometry, gap/overlap, missing/duplicate/reversed mapping, output/group count mismatch, and read-only object/Tensor ownership.
|
||
- Terminal Merge is classified as one shared physical group with no external decode boundary. Unit tests cover the valid final logical pair and terminal flag mismatch. Second Pass metadata cross-checks `group_id`, `logical_chunks`, `physical_frames`, `trim_prefix_frames`, and `terminal_merged`, with exact and three mismatch cases.
|
||
- GPU baseline and Phase6 runs used V3.8 T2VA, 3x5 seconds, 576x576, fixed seed `3801601`, Balanced 22, Audio Continuity ON, Detailed, Run Storage Off, and 4 steps. API time was 171.193 / 169.188 seconds. Both Boundary 1→2 and 2→3 reported gap 0, context 22f / video 7T / audio 37T, and PASS.
|
||
- All three raw Video latent SHA-256 values, all three raw Audio latent SHA-256 values, seed lines, and normalized Assembly Plan matched the pre-Phase6 runtime exactly. Boundary advisory count was zero, so Status diagnostics are the only expected difference. The minimum required normal 3-chunk GPU Gate passed; the optional extra Terminal Merge GPU rerun was not repeated because Terminal behavior is read-only/unit-covered and the existing G4R Production Terminal Gate remains accepted.
|
||
- Full source validation passed 920 tests. Manifest is 224/224. Runtime pre-change snapshot: `pre-v38-phase6-boundary-runtime-20260901_1240`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-chunk-boundary-gpu-gate-20260901`.
|
||
- Final combined validation passed `tools/validate.ps1`, 920 pytest tests, compileall, runtime verifier, 54 public V3.7/V3.8 workflow/UI tests, JavaScript syntax 7/7, Manifest 224/224, nine source/runtime SHA pairs, Phase5 protected-file parity 267/267, and `git diff --check`. Only the inherited `.pytest_cache` ACL notice and known `pynvml` FutureWarning remain.
|
||
|
||
## V3.8 Phase 1-6 accepted baseline / Advanced A0 PASS (2026-09-01)
|
||
|
||
- Phase 1-6 is the formally accepted V3.8 reliability baseline. Snapshot `v38-phase1-6-accepted-baseline-pre-a0-20260901_125719-0a7c670` contains 755 files, excluding only `.git` and the inherited ACL-inaccessible `.pytest_cache`; missing, extra, and SHA-256 mismatch counts are all zero.
|
||
- Three optional GPU checks remain explicitly unverified: an additional Phase 5/6 Terminal Merge GPU run, a dedicated Reference Image Cache Cold/Warm GPU A/B, and a dedicated Reference Audio Cache Cold/Warm GPU A/B. They do not block the accepted baseline. Reference Audio must return to a dedicated canary if its VAE or device path changes.
|
||
- Advanced Phase A0 Baseline Freeze / Capability Gate is formally PASS. The existing Second Pass, RefineSchedule, Refine Context, Hi-Res, and masked-transport suite passed 74 tests. Six Second Pass/Refine source-runtime file pairs match by SHA-256.
|
||
- The active ComfyUI Core is `0.34.2`. A direct CPU capability probe confirmed exact NestedTensor packing/unpacking for representative H3 Video and Audio streams: Video uses random noise with mask 1, while Audio uses zero noise with mask 0. Core preserves the separate stream shapes through mask expansion and packed sampling handoff.
|
||
- The current Second Pass contract remains Video refinement with first-pass Audio returned as the same bit-exact passthrough objects. RefineSchedule retains External, Full, Tail, and Partial modes. A0 changed no Production code, Frontend, public schema, workflow, or runtime file.
|
||
- Phase A1 is specification audit only. Before any implementation, it must define Video Only, Audio Only, and Video + Audio in terms of noise, mask, adopted output, seed namespace, and Second Pass contract identity. Its first Gate is bit-exact preservation of the existing Video Only behavior.
|
||
|
||
## V3.8 Advanced Phase A1 Refine Target / Video Only parity PASS (2026-09-01)
|
||
|
||
- Added the versioned, internal-only `refine_target_contract v1` in `v3/refine_target.py`. It freezes Noise, Mask, adopted Output, and independent Seed namespace policies for Video Only, Audio Only, and Video + Audio. RefineSchedule v1 remains the unchanged sigma-range authority, and `second_pass_contract.version=1` remains unchanged.
|
||
- Added `sample_targeted_refine_chunk()` and `run_targeted_second_pass_groups()` as separate internal paths. The accepted `sample_refine_chunk()` and `run_second_pass_groups()` implementations were not modified. Phase A1 executes Video Only only; Audio Only and Video + Audio are contract-defined but fail before Sampling as not yet implemented.
|
||
- Video Only delegates to the accepted V3.5 path. Focused parity confirms exact Video latent SHA-256, Audio LATENT object identity, refine seed, effective SIGMAS object identity, physical-group order, existing Assembly Plan fields, status, and Sampling count. The V3.5 public schema is unchanged and has no Refine Target socket.
|
||
- Validation passed 9 dedicated tests, 44 related Second Pass/RefineSchedule tests, and 929 full tests. Manifest contains 228 entries. The accepted pre-change snapshot is `pre-v38-a1-refine-target-video-only-20260901_132227-0a7c670` with 755 exact files and zero missing, extra, or SHA-256 mismatch.
|
||
- No Production facade, Frontend, workflow, public schema, active runtime file, GPU Sampling path, package version, commit, push, tag, or release was changed. Phase A1 stops at Video Only parity; Audio Only / Video + Audio execution belongs to separately approved Phase A2.
|
||
|
||
## V3.8 Advanced Phase A2 Audio Only static Gate (2026-09-01)
|
||
|
||
- Audio Only is implemented only in the internal target-aware path. It sends Video zero noise/mask 0 and Audio seeded-random noise/mask 1 to Core, adopts sampled Audio, and unconditionally discards temporary sampled Video. The returned Video LATENT entries are the original first-pass objects, guaranteeing object identity and bit-exact SHA preservation independently of current Core mask behavior.
|
||
- Audio seeds use the `h3-continuum-refine-audio-v1` namespace per physical group. RefineSchedule v1 remains unchanged and supplies only the effective sigma range. `refine_target_contract v1` remains the stream-policy authority, and `second_pass_contract.version=1` is preserved.
|
||
- Terminal Merge remains one physical group: the 3-logical / 2-physical mocked Gate samples Chunk 1 once and the terminal Chunks 2-3 pair once. Sampled Audio shape, finite values, target seeds, SIGMAS object identity, group order, and all pre-existing Assembly Plan fields pass.
|
||
- Focused A1+A2 validation passed 13 tests; the related Second Pass/RefineSchedule suite passed 48; full regression passed 933. Manifest is 229/229, compileall/runtime verifier and `git diff --check` pass. Only the inherited `.pytest_cache` ACL notice and known `pynvml` FutureWarning remain.
|
||
- Accepted snapshot: `pre-v38-a2-audio-only-static-gate-excluding-pytest-cache-20260901_134820-0a7c670`, 768 files, zero missing/extra/SHA-256 mismatch. The standard snapshot attempt stopped only on the known inaccessible `.pytest_cache` before any source edit.
|
||
- **Phase A2 CPU/static = Experimental PASS / Production HOLD.** ComfyUI Core H3 denoise-mask PR [#15988](https://github.com/Comfy-Org/ComfyUI/pull/15988) remains the upstream blocker. No GPU canary, runtime synchronization, public node/socket, Hi-Res Fix, Frontend, Run Storage, workflow, version, commit, push, tag, or release change was made. Video + Audio execution remains unavailable.
|
||
|
||
## V3.8 Advanced Phase A2 Audio Only GPU Canary (2026-09-01)
|
||
|
||
- Accepted pre-GPU snapshots are source `pre-v38-a2-audio-only-gpu-canary-source-excluding-pytest-cache-20260901_141200-0a7c670` (771 exact files) and ComfyUI_W runtime `pre-v38-a2-audio-only-gpu-canary-runtime-w-excluding-pytest-cache-20260901_141201` (387 exact files), both with zero missing, extra, or SHA mismatch.
|
||
- Synchronized only `v3/refine_target.py`, `v3/targeted_refine_sampling.py`, and `v3/targeted_second_pass.py` to ComfyUI_W; all three source/runtime SHA-256 pairs are exact. A temporary diagnostic node was loaded from an external 8190-only path and was not added to public mappings, schema, workflow, Frontend, or Production nodes.
|
||
- G1 1x5-second Audio Only Tail 6 passed: one physical group, Video object/tensor identity and SHA exact, Audio SHA changed with shape preserved, 120 frames / 5.000 seconds / 24 fps / 32 kHz stereo, and finite non-collapsed PCM.
|
||
- G2 3x5-second Balanced 22 passed: physical groups `[1] + [2] + [3]`, all three Video SHAs exact, all three Audio SHAs changed, and both 5/10-second boundaries passed the finite/non-silent/click advisory and duration checks.
|
||
- G3 FL2VA Terminal Merge passed: physical groups `[1] + [2,3]`; Chunks 2-3 remained one terminal Audio-refine sample, both Video groups were exact object/tensor passthrough, both Audio groups changed, and no partial terminal pair occurred.
|
||
- API times were 132.653 / 359.819 / 444.114 seconds. Peak device use was 15.448 / 15.537 / 15.676 GiB; 16 GiB passes with small headroom. The maximum Continuum status-reported backend RSS was 41,465.5 MiB. No OOM, NaN/Inf, decode, assembly, sampling, or duration error occurred.
|
||
- Post-Gate validation passed 933 tests, Manifest 229/229, compileall, runtime verifier, A2 source/runtime SHA 3/3, Python/JavaScript syntax, and `git diff --check`. The dedicated 8190 backend PID 19084 was stopped after Queue empty; 8188 and the frontend were not operated.
|
||
- **A2 = GPU Experimental PASS / Production HOLD.** PR [#15988](https://github.com/Comfy-Org/ComfyUI/pull/15988) remains the Production blocker. A3 Selective Refine/public node was not implemented. No commit, push, tag, or release was performed. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a2-audio-only-gpu-canary-20260901_141853`.
|
||
|
||
## V3.8 Advanced Phase A3a Video + Audio internal Gate (2026-09-01)
|
||
|
||
- Added Video + Audio execution only to the internal target-aware path. One target-derived physical-group seed in namespace `h3-continuum-refine-av-v1` creates seeded-random Video and Audio noise; both minimal masks are 1, and both sampled outputs are adopted.
|
||
- The sampling implementation shares only the Core handoff with Audio Only. Video Only still delegates unchanged to the accepted V3.5 path, and Audio Only retains zero Video noise/mask 0, sampled-Audio adoption, and original Video object passthrough.
|
||
- Physical groups remain authoritative. The 3-logical / 2-physical static Gate samples Chunk 1 once and terminal Chunks 2-3 once, adopts both finite streams, preserves Video/Audio shapes, group order, RefineSchedule SIGMAS identity, target-derived seeds, existing Assembly Plan fields, and `second_pass_contract.version=1`.
|
||
- Invalid sampled Video/Audio geometry and NaN/Inf in either adopted stream fail before output commit. Dedicated A3a tests also freeze both random-noise calls, mask geometry/value, exact SIGMAS handoff, output adoption, Terminal atomic grouping, and `refine_target_contract v1` identity.
|
||
- Focused A1/A2/A3a passed 19 tests; related Second Pass/RefineSchedule passed 62; full validation passed 939. Manifest is 230/230, compileall and runtime verifier pass, and the registered public node set remains unchanged.
|
||
- Accepted pre-change snapshot: `pre-v38-a3a-video-audio-internal-excluding-pytest-cache-20260901_144711-0a7c670`, 801 files with zero missing, extra, or SHA mismatch. The standard snapshot stopped only on the inherited `.pytest_cache` ACL before source edits.
|
||
- **A3a Video + Audio internal = CPU/static Experimental PASS.** GPU behavior is not yet claimed. A3b public Selective Refine node, Frontend, workflows, active runtime sync, A3c GPU Matrix, commit, push, tag, and release were not performed. Audio Only remains `GPU Experimental PASS / Production HOLD`; Video Only remains the Production-compatible parity candidate.
|
||
|
||
## V3.8 Advanced Phase A3b Selective Refine public node (2026-09-01)
|
||
|
||
- Added the public node ID `H3ContinuumSelectiveSecondPassExperimental`, displayed as `H3 Continuum Selective Second Pass (Experimental)` in `MiniMax H3/Continuum/Advanced`. Its fixed `Refine Target` choices are `Video Only`, `Audio Only`, and `Video + Audio`, with `Video Only` as the default.
|
||
- The node is an adapter only. Public labels resolve explicitly to the existing versioned target modes before calling `run_targeted_second_pass_groups()`. Unknown labels fail instead of falling back. No Sampling, RefineSchedule v1, `second_pass_contract.version=1`, Assembly Plan schema, V3.5 node, V3.8 Sampler schema, Frontend, Hi-Res Fix, Run Storage, or workflow behavior changed.
|
||
- Video Only remains the exact V3.5-compatible path with first-pass Audio passthrough. Audio Only remains `GPU Experimental PASS / Production HOLD`; Video + Audio remains CPU/static Experimental with GPU unverified. Status text identifies the selected target and those limitations without reading, copying, or hashing tensors.
|
||
- Registration/schema/dispatch/parity/target-output/Terminal atomic tests passed `10`; the combined Second Pass/RefineSchedule suite passed `72`; full source regression passed `949`. Compileall, JavaScript syntax `7/7`, source runtime verifier, and `git diff --check` pass. The distribution Manifest contains 231 verified entries.
|
||
- Accepted pre-change snapshot: `pre-v38-a3b-selective-refine-excluding-pytest-cache-20260901_151920-0a7c670`, 804 files with zero SHA-256 mismatch, excluding only `.git` and the inherited inaccessible `.pytest_cache`. The standard snapshot helper stopped only at that known ACL.
|
||
- **A3b Selective Refine public node = CPU/static Experimental PASS.** No runtime synchronization, backend/frontend operation, GPU run, A3c GPU Matrix, package-version change, commit, push, tag, or release was performed.
|
||
- The required `H3情報チェック` handoff formally accepted A3b with no mandatory correction. Classification remains: Video Only CPU/static parity PASS pending the A3c V3.5 GPU comparison; Audio Only GPU Experimental PASS / Production HOLD; Video + Audio CPU/static Experimental PASS / GPU unverified; the Selective Refine node as a whole CPU/static Experimental PASS.
|
||
- A3c must start from this accepted source without additional Sampling changes: snapshot the 949-pass state, minimally synchronize the new node and required target modules to ComfyUI_W, verify source/runtime hashes, restart only the backend needed for registration, freeze all generation inputs, and stop immediately if G1 Video Only differs from V3.5. A3c was not started by A3b.
|
||
|
||
## Issue #13 R2 long-form continuation re-investigation (2026-09-01)
|
||
|
||
- Formal phase state remains **A4 GPU Experimental PASS / A5 not started**. Issue #13 R2 was run as an independent diagnostic before A5; no Production Sampling, Masked AV, Assembly, Run Storage, public schema, Frontend, workflow, or package-version file changed.
|
||
- Controlled G1 reproduced a related gradual visual drift with Standard / Masked AV, Balanced 22, Audio Continuity ON, same-text List, and `6 x 10 sec`: Chunk 1 to 6 luma std `+0.028048`, saturation p95 `+0.090196`, contrast p95-p05 `+0.072273`, and Laplacian sharpness `+74.9052`. The effect is milder than the reporter's close-up example and does not establish a universal model defect.
|
||
- Matched G2 at `5 x 5 sec` did not show the same rise and is decoded-output identical to the accepted 2026-08-27 R1 Standard control: RGB24 MD5 `1e655f35cfb7b93082831d275e3ee6b5`; float32 32 kHz stereo PCM MD5 `5da19da4744431f49701bb1ba818f157` for both old and new files.
|
||
- Final protected Video/Audio prefixes were bit-exact after every physical group: G1 `5/5`, G2 `4/4`, maximum difference `0`. Sampling-internal floating epsilon is removed by the existing scoped restore. Prefix copy/overwrite, Assembly, Seam, Run Storage, and prompt-format defects are rejected for this reproduction.
|
||
- Current classification is **controlled related reproduction under long-form Standard; exact cause unresolved**. Duration/trajectory-dependent continuation feedback is a supported hypothesis, not a confirmed root cause. The earlier general claims that Reference Context alone causes the issue or Standard/Balanced 22 generally avoids it are withdrawn.
|
||
- `H3情報チェック` accepted R2 completion at G1/G2 and required no additional GPU case or Production fix. The next meaningful reproduction is R3 using the reporter's exact workflow JSON plus model/precision, seed, resolution, steps, LoRA, Spectrum/Sol, attention, Seam, and source-image details. Optional Audio Continuity OFF isolation is not an R2 completion requirement.
|
||
- Do not merge this issue into A5 Reference Video Influence/Fade; Reference Video is OFF in R2 and A5 is a separate contract. A5 remains HOLD until the R2 record is closed and the reporter-specific R3 direction is decided.
|
||
- Final validation passed `973` tests, Manifest `234/234`, runtime verifier, Python/JavaScript syntax, A4 Production source/runtime SHA `4/4`, diagnostic source/runtime SHA `1/1`, and `git diff --check`. Accepted snapshot: `issue13-r2-diagnostic-complete-excluding-pytest-cache-20260901_194903-0a7c670`.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\issue13-r2-20260901_192500\ISSUE_13_R2_REINVESTIGATION_REPORT.md`.
|
||
## V3.8 Advanced Phase A7a Packed-row Observability CPU/static Gate (2026-09-02)
|
||
|
||
- A7 readiness is **PASS** for a read-only first stage only: actual ComfyUI Core `PackedLayout` row observability and deterministic physical-group planning. A7b execution/action policy remains HOLD.
|
||
- `v3/packed_row_planner.py` accounts the immutable Core `seq_len/segments` partition as text, visual/audio conditioning, target Video/Audio, and unknown rows. It reports the existing physical-group order, `max_concurrent_physical_groups=1`, Terminal atomicity, and the earliest largest observed group as a peak-row candidate.
|
||
- Reports are Detailed-only and advisory. They do not predict VRAM/time, add a public widget, change physical grouping/order, split Terminal Merge, mutate layout/conditioning/Assembly/Run Storage/Seed/SIGMAS, synchronize CUDA, clear caches, reset allocator peaks, or stop Sampling when diagnostics are unavailable.
|
||
- Point-in-time system/CUDA/Core-free counters are observations only. Output parity and real resource behavior still require the dedicated A7 GPU Matrix.
|
||
- Validation passed A7 focused `13`, related `72`, full pytest `1017`, Manifest `238/238`, runtime verifier, Python syntax, JavaScript `7/7`, and `git diff --check` (existing line-ending warnings only).
|
||
- `H3情報チェック` formally accepted **A7a Packed-row Observability = CPU/static Experimental PASS / GPU Gate ready** with no mandatory correction. `peak-row candidate` is the required term because maximum actual rows are not yet claimed to equal the real Peak VRAM group. G4/G5 candidate changes are observational and absence of a change is not a failure by itself.
|
||
- Pre-change snapshot: `pre-v38-a7a-packed-row-observability-excluding-cache-20260902_131424-0a7c670` (`300` files, mismatch zero; excludes `.git`, inherited inaccessible `.pytest_cache`, `__pycache__`, and `.pyc`). Runtime synchronization, backend operation, GPU Sampling, commit, push, tag, release, and A7b were not performed. A5 remains HOLD.
|
||
|
||
## V3.8 Advanced Phase A7a Packed-row Observability GPU Matrix (2026-09-02)
|
||
|
||
- **A7a Packed-row Observability = GPU Experimental PASS.** Actual Core `PackedLayout` rows were observed read-only, the earliest physical group with the largest observed `seq_len` was selected deterministically as the `peak-row candidate`, `max_concurrent_physical_groups=1` remained true, Terminal Merge atomicity was preserved, and Sampling/output parity remained exact. Memory values are observations only, not VRAM/time predictions. There is no execution policy or public widget; Production behavior is unchanged. A7b and A5 remain HOLD.
|
||
- G1 `1x5` T2VA observed candidate 1 / `12,436` rows. Same-runtime Basic versus Detailed matched Video/Audio latent SHA, decoded RGB24/PCM, and Assembly semantic identity exactly; only Detailed exposed the planner report.
|
||
- G2 `3x5` continuation observed physical order `[1,2,3]`, candidate 2 / `14,112` rows, and exact per-group Video/Audio latent plus Assembly parity against the accepted A6 pre-A7 First Pass baseline. G3 FL2VA observed groups `[1]+[2,3]`, candidate 2 / `26,836` rows, exact latent/Assembly parity, and one unsplit atomic Terminal group.
|
||
- G4 three Reference Images observed candidate 1 / `14,499` rows with `1,008` condition-video rows. G5 124-frame Video Guide at Efficient 0.4 MP observed candidate 1 / `29,706` rows with `14,800` condition-video rows. Both same-runtime Basic/Detailed comparisons retained exact latent, decoded media, and Assembly parity.
|
||
- Every Detailed group satisfied segment-row sum equals `seq_len`, exactly one target Video and Audio segment, target rows greater than zero, fixed execution order, and earliest-largest candidate selection. No OOM, NaN/Inf, Decode, Assembly, duration, or PCM error occurred. Peak observed VRAM was `15.531 GiB`, peak backend RSS `47.881 GiB`, and minimum available RAM `4.301 GiB`; the 16 GiB Gate passes with small VRAM headroom.
|
||
- The initial G1 evidence runner STOP was caused only by comparing the existing volatile Assembly field `_runtime_started_at` in a full JSON hash. Product latents and decoded outputs were already exact. The evidence runner alone was aligned with the established semantic-identity rule and resumed without re-Sampling; no Production fix was made.
|
||
- Post-GPU validation passed full pytest `1017`, Manifest `238/238`, runtime verifier, Python syntax, JavaScript `7/7`, source/runtime SHA `3/3`, and `git diff --check` with only inherited line-ending warnings. Dedicated 8190 was stopped after Queue `0/0`; 8188 and Frontend were not operated. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a7a-packed-row-gpu-matrix-20260902_133600\summary.json`.
|
||
- Submitted the complete Gate to pinned `H3情報チェック` and incorporated its formal approval. It requires no correction, additional GPU Gate, or re-Sampling; it accepts the reused A6 pre-A7 G2/G3 baseline because latent SHA, Assembly semantic identity, order, and decoded health all pass. A7a is formally closed; A7b remains a separately approved future action-policy phase.
|
||
- Final accepted source snapshot: `v38-a7a-gpu-experimental-pass-accepted-final-excluding-cache-20260902_141417-0a7c670` (`303` files, mismatch zero; excludes `.git`, `.pytest_cache`, `__pycache__`, and `.pyc`).
|
||
|
||
## V3.8 Advanced Phase A8a Adaptive Continuity Observer (2026-09-02)
|
||
|
||
- **A8a Adaptive Continuity Observer = GPU Experimental PASS.** The observer emits a deterministic, JSON-safe decision trace after Production facts are resolved. Every record fixes `execution_applied=false`; it cannot alter context frames, continuation transport, masks, latent tensors, conditioning, Seed, SIGMAS, Sampling, physical grouping, Assembly, State, Session, or Run Storage identity/schema.
|
||
- The versioned pure planner is `v3/adaptive_continuity.py`; `v2/sequence.py` exposes it only in Detailed status reporting. Reused Run Storage prefixes are reconstructed from saved entry metadata, and FL2VA Terminal Chunks 2-3 remain one atomic observation. The trace is advisory and makes no Issue #13 mitigation claim.
|
||
- GPU G1 `1x5`, G2 `3x5`, Issue #13 G3 `6x5 same-prompt`, G4 FL2VA Terminal `[1]+[2,3]`, and G5 Run Storage restart/Review replay all passed. Accepted latent, normalized Assembly, decoded RGB/PCM, prompt/seed/SIGMAS signatures, prefix reuse, Last Frame/Terminal behavior, and restart recovery remained exact where applicable. G5 progressed `review_ready Chunk 1 -> review_ready Chunk 2 -> complete Chunk 3`; the Sampler executed after restart while accepted prefixes were reused from storage rather than ComfyUI cache.
|
||
- Overall peak observed VRAM was `15.588 GiB`; maximum status-reported backend RSS was `46.478 GiB`; minimum available RAM was `2.979 GiB`. The 16 GiB Gate passes with small headroom. No OOM, NaN/Inf, Decode, Assembly, duration, PCM, Terminal partial-pair, or replay failure occurred.
|
||
- Post-GPU validation passed full pytest `1041`, Manifest `242/242`, runtime verifier, Python syntax/compileall, JavaScript `7/7`, source/runtime SHA `2/2`, and `git diff --check` with inherited line-ending warnings only. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-a8a-adaptive-continuity-gpu-matrix-20260902_180928\A8A_ADAPTIVE_CONTINUITY_GPU_PASS_REPORT.md`.
|
||
- The pinned `H3情報チェック` task formally accepted A8a with no mandatory fix, extra GPU Gate, or re-Sampling. **A8b execution policy remains HOLD; A5 remains HOLD.** Sage+Sol/Spectrum matrix work is required only if A8b execution is later considered for promotion.
|
||
- Production behavior, A7b, public schema, workflows, Frontend, package version, Run Storage/State/Session/Assembly contracts, and external wrappers remain unchanged. Dedicated 8190 was stopped after Queue empty; 8188 and Frontend were not operated. No commit, push, tag, or release was performed.
|
||
- Final accepted source snapshot: `v38-a8a-gpu-experimental-pass-accepted-final-excluding-cache-20260902_190700-0a7c670` (`309` files, mismatch zero; excludes `.git` and ignored cache/bytecode files).
|
||
|
||
## V3.8 Release / Registry Package Staging Re-Gate (2026-09-02)
|
||
|
||
- **Package Staging Re-Gate = PASS / Release & Registry readiness = GO.** The previous package-only blockers are resolved. Public Surface Cleanup remains PASS/Close; Development Track remains independent and available. A5/A8b remain HOLD and P1a/P2 remain Deferred.
|
||
- `.comfyignore` excludes `tools/*` and restores only `tools/__init__.py` plus the Production-required `tools/p20_file_backed_tensor_poc.py`; it excludes all historical examples and restores only `examples/workflows/MiniMax_H3_Continuum_V38.json`. `AGENTS.md`, `PROJECT_STATE.md`, `WORKLOG.md`, tests, and source `MANIFEST.sha256` are excluded from the Registry archive.
|
||
- The development/source Manifest was preserved and extended for the two required distribution files: `247/247`, missing 0, mismatch 0. The staged archive carries a separate non-self-hashing `REGISTRY_MANIFEST.sha256`: `138/138`, missing 0, extra 0, mismatch 0.
|
||
- Official `comfy-cli==1.18.0` validate and pack exited 0. The final archive contains 139 files, source/stage SHA is `139/139`, isolated import passes, and public export/display are `6/6`. Tools are exactly the two allowlisted files; examples are exactly the V3.8 workflow; historical assets, internal development documents, source Manifest, and tests are all 0.
|
||
- Focused regression passed 6; full regression passed 1047 using a dedicated writable basetemp. The CLI's existing future-compatibility E702 warnings remain non-blocking and were not changed in this packaging-only scope. `git diff --check` exits 0 with inherited line-ending warnings only.
|
||
- Final archive SHA-256 is `52115DA30B55F91699D01352580CE3A76F462EF27701C31183201597568A50EB`. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-package-staging-regate-20260902_224219\PACKAGE_STAGING_REGATE_REPORT.md`. Pre-change snapshot: `pre-v38-release-package-fix-filtered-20260902_224003-0a7c670` (`312/312`, mismatch 0); the standard helper stopped only on the inherited `.pytest_cache` ACL denial before edits.
|
||
- The designated `H3情報チェック` task formally confirmed no mandatory blocker and accepted **Release / Registry readiness = GO**. Production/runtime/backend/frontend/GPU Sampling were unchanged. No commit, push, tag, Registry publish, or release was performed; those remain separate explicit-approval actions.
|
||
|
||
## V3.8 Release Hygiene (2026-09-03)
|
||
|
||
- Current release-hygiene records cover the V3.8 seven-node surface, AUDIO-R1, N2b/N2c, O1, and the Size Source UX. README files distinguish the V3.8 standard path from historical V3.7/Hi-Res notes; `INSTALL_JA.txt` now documents V3.8 installation and the seven public nodes.
|
||
- Development `MANIFEST.sha256` is `253/253` and Registry `REGISTRY_MANIFEST.sha256` is `102/102`, both with zero missing, mismatch, or duplicate entries. The Registry candidate has zero tests, exactly two tools modules, and exactly one V3.8 workflow.
|
||
- A 102-file manifest-isolated import from `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-release-hygiene-isolated` loaded only staged code and verified public display/export `7/7`. Runtime verifier passed; tracked Markdown relative links and PNG signatures are clean.
|
||
- Full source pytest baseline is `1117 passed, 0 failed`; JavaScript and compileall pass. **Release / Registry readiness remains HOLD**: this hygiene check did not run official comfy-cli validate/pack or create a release archive. Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-release-hygiene-20260903_201500\PACKAGE_RELEASE_HYGIENE_REPORT.md`.
|
||
|
||
## V3.8 Run Storage Windows Atomic Replace Reliability PASS (2026-09-03)
|
||
|
||
- `run_storage._write_json()` retains complete-temp → atomic `os.replace()` → canonical-manifest semantics. On Windows only, access-denied/sharing-violation errors receive bounded retries of 20/50/100/200/400 ms; no delete-and-rename or non-atomic fallback exists.
|
||
- Retry exhaustion raises a diagnostic `RunStorageError` while preserving the existing manifest and cleaning the temporary candidate when possible. Schema v3, Sampling v5, State, Session, Assembly, Branch Provenance, canonical-head semantics, Take identity, Seed, and SIGMAS are unchanged.
|
||
- The prior failure reproduced before the fix in a fresh native Windows temp root. Post-fix: atomic transient/permanent cases and the corrupted-newest-branch fallback pass; the fallback created 20 fresh test roots without an iteration failure; Run Storage/Branch Provenance suite passed `78`; full pytest passed `1120` with zero failures.
|
||
- Development Manifest is `253/253`, Registry Manifest is `104/104`, public verifier is `7/7`, JavaScript syntax passes, protected non-persistence contracts are unchanged, and `git diff --check` passes. Release / Registry readiness remains HOLD pending official package staging.
|
||
|
||
## V3.8 Official Package Staging Re-Gate PASS (2026-09-03)
|
||
|
||
- **RELEASE PACKAGE GATE = PASS. Release / Registry readiness = GO candidate.** Commit, push, tag, release, and Registry publication still require separate approval.
|
||
- Official isolated `comfy-cli 1.18.0` completed `comfy node validate` and `comfy node pack` with exit `0`. The CLI configuration home was process-isolated due an ACL denial in the pre-existing user CLI config directory; source and package were unaffected.
|
||
- Archive SHA-256: `B9278516D0A95745ED6EFFF7ABC693574F52E2DA9CCC2F667E3A599725E9A673`. Its 105 files are Registry payload `104/104` plus its Manifest; source ↔ stage and stage ↔ archive missing/extra/mismatch are `0`.
|
||
- Expanded-stage-only runtime verification passed Public `7/7`, native PackedLayout, and multi-reference-audio rows. Installer/verifier dependencies are self-contained; only the V3.8 workflow is packaged; forbidden artifacts, bytecode/cache, local user paths, and recognizable credentials are absent.
|
||
- Evidence: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-official-package-staging-regate-20260903_210529\PACKAGE_STAGING_REGATE_REPORT.md`. Snapshot before result recording: `pre-official-package-staging-gate-record-filtered-20260903_211009-0a7c670` (`321/321`, mismatch `0`); the ordinary snapshot helper stopped only on the inherited `.pytest_cache` ACL denial.
|
||
|
||
## V3.8 Review seed-lineage UI fix (2026-09-05)
|
||
|
||
- Live Run Storage evidence confirmed that the failed Continue attempt used Base Seeds `588314452213700` and `922887055962422`, producing distinct lineage hashes and two independent Group 1 revisions. The backend correctly rejected cross-lineage reuse; the defect was frontend seed-control handling.
|
||
- V3.8 now resolves the actual ComfyUI seed control through `base_seed.linkedWidgets`, automatically changes it to `fixed` whenever `Run = Review Each Chunk` is selected or restored, and applies the same protection to Review and Take action shortcuts. The pre-Queue guard remains as a final safety check.
|
||
- Sampling v5, Run Storage v3, State, Session, Assembly, seed derivation, Review resolver, and public schema are unchanged. Related regression passed `166`; full regression passed `1128/1128` under the ComfyUI_W venv. Runtime verifier passed Public `7/7` after source/runtime synchronization.
|
||
- The user-controlled backend was not started, stopped, or restarted and was offline during the final live-route check. The browser must reload once after the backend is next started to activate the synchronized JavaScript. Release / Registry readiness remains HOLD pending a fresh package staging Gate containing this fix.
|
||
|
||
## V3.8 Continuation Regression Final Gate A/B/C (2026-09-07)
|
||
|
||
- GPU A/B/C completed under corrected live `00058`: 896×896, 24 fps, 4×8 seconds, seed `595713724167856`, Turbo4 v1.2, Euler/simple/6, KJ Sage, Balanced22, Audio Continuity ON. A=Standard/Masked AV, B=Compatibility/Reference Context, C=Standard plus isolated R6-A Pin Renorm.
|
||
- Every Arm produced 768 frames, PCM `[1,2,1024000]`, raw Video4/Audio4 and a clean owned-backend shutdown. A/B/C Group1 raw AV is bit-exact. No OOM, crash, nonfinite data or AV contract failure occurred.
|
||
- C recorded exactly three R6-A events, but every applied gain was `1.0` because later source-tail std remained below the Group1 reference and amplification is forbidden. C is therefore a valid no-op: A/C seed, SIGMAS, noise, conditioning, Sampling input/output, raw AV, PCM and decoded AV are bit-exact. This proves GPU no-op parity, not R6-A efficacy.
|
||
- The tested seed did not reproduce Issue #13 sharpness ratchet: Laplacian decreased A `54.803→41.359` and B `54.803→40.399`. B showed greater saturation and skin-highlight growth and its 8-second PCM edge exceeded the pre-set A-relative screen on both channels. Visual/audio human review remains recommended, but there is no evidence here to replace Standard with Compatibility.
|
||
- Current launch direction: keep Standard/Masked AV as default; do not promote Compatibility from this result; retain R6-A as Experimental/Default OFF pending active-gain multi-seed or reporter-exact validation. Issue #13 is not Solved. Production/runtime/public schema remain unchanged; no commit, push, tag or release was performed.
|
||
- Evidence and report: `D:\Codex\_test_results\ComfyUI-H3-Continuum\v38-continuation-regression-final-gate-20260907\FINAL_ABC_REPORT_JA.md`.
|
||
|
||
## V3.8 Review Each Chunk live 3 x 5s acceptance PASS (2026-09-08)
|
||
|
||
- The remaining real-screen acceptance is now **PASS** on the user-managed `127.0.0.6:8188` backend. A single 3 x 5-second Review Each Chunk run advanced through `Chunk 1 is ready for review`, `Chunk 2 is ready for review`, then `Saved sequence is complete` only after Chunk 3. Chunk 1 and Chunk 2 each exposed `Use it and continue`, `Try this chunk again`, and `Use it and finish the rest`; the final state correctly removed no-op continuation controls.
|
||
- The prior CPU-only frontend test had exercised an unsupported `beforeQueuePrompt` hook. ComfyUI frontend `1.49.6` actually runs widget `beforeQueued`, serializes through `serializeValue`, then emits `executed` and `execution_success`. `web/project_id.js` now uses that supported lifecycle to queue the transient Chunk 1 boundary without mutating visible/saved widgets and clears setup only for the exact node, Run name, setup token, and prompt identity.
|
||
- Live prompts `8a9a2275-7f31-4e3b-bf7c-95747feefc2b`, `5427b850-0490-4b07-8243-77f737be7ba3`, and `ef38c575-faa8-45ea-8485-6ced1500298e` all completed successfully. Each reported exactly one new physical group: `1/3`, `2/3`, and `3/3` logical chunks available; Run Storage progression was `0 reused + 1 generated`, `1 + 1`, then `2 + 1`. Sampler node `312` was not returned as cached in any prompt; only upstream nodes reused ComfyUI cache.
|
||
- Focused supported-lifecycle regression passed `1/1`; full CPU regression passed `1156/1156` in `50.60s`; JavaScript syntax passed. Source and installed runtime `web/project_id.js` are byte-identical at SHA-256 `78BA840BB4B57D21BC17AAB9F5CA3C89E3BC3C7AAE1A027BBD3FF54512717A53`.
|
||
- Production Python, Sampling, transport, backend lifecycle, and workflow JSON were unchanged. The user performed every GPU Queue/Continue action; Codex did not start GPU work or restart/stop the backend. No commit, push, tag, release, or Registry publication was performed.
|
||
|
||
### Newly exposed Render History defect
|
||
|
||
- Opening Render History after the accepted run exposed a separate existing bug. The project index contains `76` group revisions across `59` sampling lineages, while the current canonical lineage contains only the three accepted revisions. The frontend presents all project revisions as selectable Takes; the backend correctly validates only the current contract lineage.
|
||
- Selecting old Group 2 revisions `9db2a9287ef18796b770db00` or `f968bd1131c20912e207de0a` and queuing an action fails safely with `selected Take provenance is invalid: group revision parent is missing`. The message refers to absence from the current-lineage validation catalog, not deletion of the accepted canonical chain.
|
||
- Repaired in `web/project_id.js`: the actionable catalog is now restricted to the canonical sampling lineage and revisions whose complete contiguous ancestry resolves inside that lineage. A stale hidden selection falls back to the corresponding canonical revision and resets its pending Take action. If settings differ from the reviewed result, Take application controls are disabled while read-only browsing remains available. Backend validation remains unchanged.
|
||
- The history body is now a compact summary capped at 124 px. Its custom draw uses the reserved computed height even when the frontend supplies a single-row draw hint, eliminating the large blank panel and the one-line-only rendering observed live.
|
||
- Mixed-lineage/orphan, selection fallback, compact rendering, and existing Review/Take regressions passed `59/59`; full CPU regression passed `1156/1156` in `52.09s`; JavaScript syntax and both manifests passed. Source/runtime JS SHA-256 is `549A8150BC926BF142758B1610E3730547D6ABCFBCDDB0B37ED8BAAAF86B38CB` after one-file deployment with Queue `0/0`. The current browser tab remains unreloaded because its title carries an unsaved marker; live visual confirmation requires saving or intentionally discarding that tab's changes, then reloading.
|