Archive fork in favor of upstream and DUMAS nodes
tests / test (push) Canceled after 0s
tests / windows-storage (push) Canceled after 0s

This commit is contained in:
2026-10-20 10:46:16 +00:00
parent 881dcbcfba
commit 3716b02a57
76 changed files with 12140 additions and 4410 deletions
+128 -25
View File
@@ -1,35 +1,110 @@
# Public development log
## 2026-10-19 — Second Pass tiled sigma precedence
## 3.9.9 — KJ Preview Override / Review continuation
- Fix `H3ContinuumSecondPassV35` so the tiled path actually forwards the
connected SIGMAS schedule instead of silently rebuilding it from the node's
`upscale_steps` value.
- Give a dedicated `upscale_sigmas` connection priority over the main SIGMAS
connection, preserve widget fallback behavior for the other tiled controls,
and report the selected sigma source for diagnosis.
- Add focused node-level regression coverage for both sigma precedence cases.
Fix new-Run v6 read-only with the separately audited upstream KJ profile, and
show saved reasons when reuse is unavailable. Keep the original local profile,
Sampling v6 guards, old Takes and official workflow bytes.
## 2026-10-15 — CSGlide-style timeline controls
CPU: **2,075 passed / 2 skipped / 2 known expected failures**, plus overlapping
90-case JS coverage. Upstream KJ GPU and normal-ComfyUI browser Continue both
passed: **Chunk 1 preserved, Chunk 2 generated, one reused / one new**; all four
AV outputs fully decoded. Browser checks used the existing local KJ profile.
These are targeted repair checks, not general quality or all-provider acceptance.
- Match CSGlide's shot-timeline interaction: one prompt editor per clip,
coloured shot bands, and boundary sliders that move adjacent shot edges
together.
- Add `@ Add shot` plus a line-start `@` keyboard shortcut for rapid prompt
authoring without changing the emitted Timeline format.
- Make `Default chunk duration` the only duration control; all clip tabs now
share it and legacy per-clip duration values are ignored.
- Ensure the final shot always ends at the full chunk duration; adding a shot
splits the previous final shot instead of leaving a grey tail.
Restart ComfyUI and hard-refresh after updating. On Preview Override, disconnect
optional VAE/Audio VAE and set Tiny VAE `none`; retain generation VAEs. Correct
unsupported settings before starting a new Run. Old unsafe/v5 Takes remain
readable and are not automatically made reusable. No workflow rewiring or Take
deletion is required. [Full release notes](docs/RELEASE_NOTES_V399.md).
## 2026-10-15 — Sequence prompt builder
- Add `DumasH3SequencePromptBuilder` and its tabbed ComfyUI editor.
- Validate clip-local slot ranges and deterministically convert them to
absolute `[start-end]` timeline blocks, preserving one STRING output for
direct wiring into Continuum.
- Add focused unit coverage for formatting, normalization, validation, and
registration/schema behavior.
## 2026-10-16 — V3.9 Turbo Sample
Add the user-supplied sample separately: Auto, 2 x 8s, multiline image-anchored
dance prompt, fixed background/camera, Thai rap and one enabled Turbo LoRA.
Preserve the supplied 6-step/Draft settings and document their unverified pairing
with the 4-step/768p LoRA. Fresh workflow/project IDs avoid user Run identity reuse;
remove misleading auto-download metadata for the saved custom model. Standard
workflows, existing Takes and runtime code stay unchanged. No new GPU validation,
runtime deployment, release or Registry publication is included.
Local CPU checks: 17 passed, zero CUDA initialization attempts. Workflow link
consistency, gallery/ZIP bytes, saved settings and both manifests were checked.
## 2026-10-16 — Japanese README organization
Align the Japanese README with the compact English entry point, including exact
model/VAE/text-encoder downloads, model folders, external node links and optional
Turbo guidance. Keep current compatibility and known limitations explicit.
Move all former Japanese detail and ten image references into USER_GUIDE_JA.md
and README_HISTORY_JA.md, preserving historical context and rebasing links.
Include these text guides in the Registry selection and update both manifests.
English documentation, product code, images, official workflows and existing
Runs/Takes remain unchanged. This is local documentation work, not a new release.
Local checks: 5 existing tests passed, 95 relative file/heading links resolved,
all ten original image references and original detailed text were retained, and
both manifests matched (464 source entries / 162 Registry entries). The Japanese
README is 247 lines instead of 801. Its 18 external destinations match English.
No CUDA initialization, runtime deployment, commit or push was performed.
## 2026-10-16 — condensed README and retained reference material
Organize the English README as a short installation and operation entry point.
Keep the version heading, fixed-version downloads, getting-started video, Review /
regeneration steps, history safety and bounded v3.9.8 compatibility/known limits.
Move the former detailed controls and screenshots into docs/USER_GUIDE.md, and
release/legacy implementation material into docs/README_HISTORY.md; retain all
original sections and image assets, with rebased links and historical context.
Follow-up: list the template-selected H3 diffusion model, text encoder, Video/Audio
VAEs with provider downloads and ComfyUI model folders. Link Spectrum, KJNodes,
rgthree and optional task-matched LightX2V Turbo LoRAs. Explain the saved INT8
Video VAE versus old FP16 download metadata without altering workflow bytes.
Include the two text references in the Registry selection and update both source
manifests. Product code, Japanese README, official workflows, stored Takes,
release tags/assets and runtime deployment are unchanged by this documentation work.
Local documentation checks: 5 existing tests passed; all 143 relative file/heading
links resolved, all 39 original image references were retained, and both manifests
matched (462 source entries / 160 Registry entries). README is 267 lines, down
from 1,574. No code, workflow, image or Japanese README changes were made.
## 2026-10-16 — published v3.9.8 GPU recheck documentation
Record a fresh GPU recheck after deploying and restarting the exact published tag `58fbd855`: PyTorch/KJ and Comfy Kitchen INT8 + audited local Spectrum/KJ each passed separate-Queue Chunk 1 to 2 continuation. KJ retained images, contracts and original Chunk 1 stayed equal, and all four AV outputs decoded. Existing Takes/workflows were hash-protected. README and Release notes distinguish this published-tag result from the earlier candidate gate and retain exact provider/no-preview-VAE/Tiny-none scope plus browser, quality and performance limits. This is a documentation-only main update; the original v3.9.8 tag and package assets remain unchanged.
## 2026-10-16 — package 3.9.8 / KJ Preview Override Phase A
- Audit the actual KJ Preview Override file, SHA-256 `d07c8e5ff7d7abd3e31acefe2cf29f618b5da5341b3b434e72ec2ac0953230b0`, rather than accepting KJNodes 1.5.2 by version alone. Pin sibling helpers and Core preview/dispatcher source, then compare loaded code/defaults/classes without importing or executing the provider during observation.
- Add a dedicated adapter only at `outer_sample/kj_preview_override`. Preserve all fixed wrapper settings and Core's key/list order. Ignore only the exact builtin transient `frames` list after the provider audit; preserve the generic no-KJ observer and all other reuse guards.
- CPU scope covers real-provider frame changes, reconstruction, fixed-setting/order changes, monkeypatch/source rejection, VAE/Tiny VAE exclusion, combined Attention/Spectrum identities, saved prefix bytes and old unsafe history. Full real-Core CPU gate: 1,995 passed / 3 skipped / 2 existing Inspector expected failures / zero unexpected failures and CUDA initialization attempts; the 48 dedicated KJ tests passed without skips. Normal ComfyUI GPU acceptance passed four Core-API Queues: PyTorch and Comfy Kitchen INT8 + the audited local Spectrum implementation, each with separate-Queue Chunk 1 to 2 continuation and held preview images. Both second Queues reused one chunk and generated one; original Chunk 1 bytes and contracts were unchanged. All four video/audio outputs fully decoded. See docs/KJ_PREVIEW_OVERRIDE_V398.md for exact support and limits. GitHub main/tag CI, Release asset integrity and Registry status are separate publication checks.
## 2026-10-15 — package 3.9.7 / linked-input regeneration cache
- Keep Full Run automatic variations reachable when Core leaves linked Chunks, Nonce, storage, regeneration boundary or generation mode unresolved during cache-key evaluation.
- Validate the actual chunk range at execution, and retain stable caching for known ordinary Full Run, storage OFF and directly specified fixed nonces. An unresolved external nonce conservatively reevaluates the node; Run Storage still owns prefix reuse.
- Keep Sampling Contract v6, Take protection, public schemas and official workflows unchanged. Add direct-input and real-Core cache regressions for both V3.8 and V3.9.
## 2026-10-15 — package 3.9.6 / regeneration intent and automatic variations
- Separate saved manual regeneration intent from the legacy stale Chunk 1 / nonce 0 state. Retain earlier unmarked later-chunk/fixed-nonce settings, and clear intent metadata when Auto or a one-shot restart consumes it. No widget order or API inputs change.
- Bypass Core output caching only for Full Run explicit regeneration with Run Storage on and automatic nonce 0. Run Storage remains the authority for compatible prefixes, interrupted variations and new completed variations; its Sampling v6 checks are unchanged.
- Add restored-workflow Queue cases for both public Samplers and a CPU-only real-Core cache lookup gate. Keep historical unsafe/v5 Takes read-only and official workflow bytes unchanged. Full CPU gate: 1,908 passed / 2 skipped / 2 known Inspector expected failures; zero unexpected failures and CUDA initialization attempts. Normal ComfyUI browser save/reload retained Chunk 2 / nonce 5 and Chunk 1 / nonce 0 after history loading. Five new-Run GPU queues passed continuation, Chunk 2 regeneration and two identical Full Run / nonce 0 submissions; earlier Chunk 1 bytes were preserved. GPU scope: PyTorch, Spectrum OFF, 512x512, 8-step Turbo; all five media outputs fully decoded. See docs/REGENERATE_INTENT_V396.md. GitHub main/tag CI, Release asset integrity and Registry status are verified separately during publication.
## 2026-10-15 — package 3.9.5 / Issue #31 publication
Main Linux/Windows CI passed. Correct the tag-presence probe to branch on the API exit status: a missing-ref error body must not be treated as an existing commit SHA. Product code and accepted GPU signatures are unchanged.
Consolidate the Attention fix and finite Spectrum binding adapter on the exact 3.9.4 public source tree. Add a separate audited upstream Spectrum 0.2.29 profile without changing existing local-provider identities. Keep old unsafe history, public node IDs, contracts and all official workflows. CPU/GPU and publication checks are recorded separately in docs/ISSUE31_SAMPLING_SIGNATURES.md.
## 2026-10-14 — Issue #31 Phase 1, unpublished
- Final standard CPU validation: 1,777 passed / 1 skipped / 2 existing Inspector expected failures; no failures. All three standard commands recorded zero CUDA initialization attempts. Native PackedLayout, prompt planning and twelve-node registration checks passed.
- Limit production edits to the MODEL attention signature adapter in run_storage.py. Use already-loaded Core registries and implementation bindings; never import Core/Kitchen or execute an attention kernel during observation.
- Include both container functions and provider implementations/state without persistent addresses. Keep unrelated v6 signatures and every existing reuse guard; do not rewrite old unsafe manifests or raw files.
- Add real-Core CPU signature, clone/recreation, separate-process/hash-seed, mutation detection, prefix reuse and history immutability regressions. Kernel registration in CPU fixtures is structural evidence, not INT8 execution acceptance.
- GPU Review continuity, Spectrum/KJ execution, browser behavior and reporter-exact reproduction remain separate. No public node, official workflow, installed runtime, package version, commit or release changes.
## 2026-10-14 — package 3.9.4 Registry packaging repair
@@ -84,3 +159,31 @@ Earlier product behavior is described in CHANGELOG.md and the historical Release
- Add 3.9.3 and retained 3.9.2 Release, custom-node ZIP and checksum links to the current English/Japanese guides. Update the installation heading to 3.9.3 and retain the implemented settings screenshot and historical validation records.
- Refresh Registry then development manifests from saved bytes. The existing release-metadata/public-surface CPU gate passes 8 tests with zero CUDA initialization attempts; all 25 tracked JSON/ZIP assets retain their bytes.
- Product code, package metadata, public schemas and saved-data contracts are unchanged. GitHub Release/asset and Registry publication are verified separately; no new GPU, long-form performance or physical SSD-write validation is claimed.
## 2026-10-15 — Issue #31 Spectrum binding identity candidate
Added a dedicated, fail-closed Spectrum binding adapter without changing the generic observer, Sampling v6, runtime signature v1, Graph v4, public nodes, workflows or package version. The immutable config fields and audited source/loaded-callable contract identify generation; reviewed runtime history, counters and profile caches do not change persistent identity. Unsupported structures, additional attributes and altered implementations remain non-reusable. Existing unsafe Takes are preserved without migration. CPU and GPU evidence are recorded separately; no commit or publication is authorized.
Issue #31 Phase 3 validation: the final standard CPU entry point passed 1,834 tests, 1 skip and 2 existing expected Inspector failures, with zero CUDA initialization attempts. The 57 new Spectrum tests cover all 27 config fields, transient-state/process stability, implementation/extra-state refusal, unaffected v6 contract identity and old unsafe Take protection; the 36 Phase 1 attention tests also pass. Under one audited local provider implementation, both PyTorch+Spectrum and Kitchen+Spectrum persisted new Chunk 1 Takes and reused them from a separate Queue to generate Chunk 2, with unchanged prefix hashes, identical safe contracts and fully decodable 10-second AV output. This is a targeted local GPU gate, not acceptance of other providers, KJ, browser behavior, subjective quality or benchmarks. No commit, push or publication occurred.
## 2026-10-20 — External Chunks Reference assignment table candidate
Only the V3.9 Reference Images helper presentation changes: external Chunks links make the destination count unresolved, expose assignment range 1–16, and label All/Off with that range. Direct counts retain the existing table. The frontend does not resolve arbitrary upstream values or alter native Chunks, saved selectors, Sampling identity or stored Takes. Added 21 JavaScript cases and four public-input CPU cases cover stale widget counts, zero/unrestored links, multiple Samplers, range actions, future assignments and callback stability. Existing CPU coverage passed 310 tests with two known Inspector expected failures; the four added CPU cases and metadata checks also pass. Targeted browser acceptance passed for external INT with stale widget values, live disconnection/restoration, actual KJ Set/Get routing, all-page All/Off actions, saving/reloading and workflow tab switching. The candidate display code was deployed to the chosen runtime and reloaded without restarting its backend. Existing saved Takes and workflows were preserved. No GPU generation, commit, push or public release occurred.
## 2026-10-20 — Spectrum source/runtime audit portability candidate
Normalize CRLF to LF in Spectrum provider-file and module-closure source checks. Add fixed Python 3.12 profiles generated from the same hash-verified 44-file and upstream 0.2.29 55-file sources. Retain the original Python 3.13 profile bytes and valid identity descriptors. Unknown code, file sets, closures and runtime extensions remain non-reusable; unsupported Python runtimes are not promoted. Old unsafe/v5 Takes remain readable without conversion, and public nodes, Sampling v6, Graph v4 and official workflows are unchanged.
The full standard CPU/JS gate passed 2,096 tests with two skips and two existing Inspector expected failures; compile and offline runtime verification passed with twelve public nodes. Targeted Spectrum/Attention coverage passed 169 tests with one legacy-profile skip. Python 3.12 component coverage passed 112 tests with one legacy-profile skip, checking all 27 configuration fields, code/state refusal and newline parity; 21 independent older-Core/Attention cases were outside that component gate. Separate-process storage checks passed 16 PyTorch/Kitchen cases on Python 3.13 and eight Spectrum/default-model cases on Python 3.12. Actual Take read-only JavaScript agreed with all 24 storage results, and all six previously valid complete Sampling Contracts remained identical. No CUDA initialization, new GPU generation, browser acceptance, commit, push or publication is claimed. The older Core/Kitchen observation limitation remains separate from Spectrum recognition.
## 2026-10-20 — 3.9.10 publication preparation
Prepare the tested Spectrum audit portability and Reference assignment display
repairs as the next 3.9 patch. Use SemVer 3.9.10 for GitHub and Registry while
reserving 4.0 for the major update. Preserve existing 3.9.9 and historical tags,
assets and official workflow bytes. Publish CPU/browser scope explicitly without
claiming new GPU acceptance. Version metadata, release workflows and package
manifests identify the same candidate; public asset and CI checks follow upload.