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
+61 -9
View File
@@ -1,15 +1,67 @@
# Changelog
## Unreleased
## 3.9.10 — Spectrum reuse and Reference assignment display (2026-10-20)
- Make Second Pass tiled-upscale schedule precedence explicit: a connected
dedicated upscale SIGMAS socket wins, followed by the existing main SIGMAS
socket, while the other tiled controls use their node widgets when their
sockets are not connected. The status report now identifies the sigma source.
- Add the DUMAS H3 Sequence Prompt Builder with clip tabs, local timeline
slots, and one canonical absolute Timeline prompt output.
- Refine its UI around the CSGlide shot-timeline pattern: one multiline shot
editor, coloured shot bands, and shared boundary sliders.
- Normalize CRLF source line endings in both Spectrum file and module-closure audits. Keep the pinned source identities and existing Python 3.13 profile bytes unchanged.
- Add separate audited Python 3.12 profiles for the supported local and upstream 0.2.29 implementations. Keep unknown source/runtime state non-reusable, preserve old unsafe/v5 Takes, and require a new Run for newly supported configurations. Sampling v6, Graph v4, public nodes and workflow bytes are unchanged. No new GPU acceptance is claimed for this repair.
## Unreleased — external Chunks Reference assignment table
- Treat externally connected Chunks as unresolved in the V3.9 Reference Images
helper. Display assignment range 1–16 with the reason and explicit All/Off
range labels; retain direct-count tables and saved future assignments.
- Keep existing selectors, backend routing, Sampling/Run Storage contracts,
public inputs and official workflows unchanged. Loading or refreshing does
not automatically expand old selections.
## 3.9.9 — upstream KJ Preview Override / Review continuation (2026-10-20)
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.
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.
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).
## Unreleased — V3.9 Turbo Sample (2026-10-16)
- Add a separate user-supplied Auto / 2 x 8s Turbo dance and Thai rap sample, gallery copy, ZIP and setup notes. Preserve standard workflows, saved Takes and product code. This source-only sample has no new GPU acceptance and does not replace v3.9.8 release assets.
## 3.9.8 — Issue #31 Preview Override Phase A (2026-10-16)
- Post-release documentation: after installing and restarting published tag `58fbd855`, both PyTorch/KJ and Kitchen + audited local Spectrum/KJ passed new-Run, separate-Queue Chunk 1 to 2 continuation with retained previews, unchanged prefix/contracts and four fully decoded AV outputs. Record the limited audited profile and unverified browser/quality/performance scope. Keep the original tag/assets and existing Takes/workflows.
- Add a finite KJ Model Preview Override signature for the explicitly audited source profile. Preserve fixed preview settings, loaded implementation identity and cross-key/list wrapper execution order; omit only the reviewed builtin `frames` list.
- Limit reuse support to VAE/Audio VAE disconnected, KJ Tiny VAE `none`, and Core NoPreviews/Auto/Latent2RGB. Unknown source, altered helpers, additional state and approximate-VAE preview remain non-reusable. Generation itself retains the existing fallback.
- Keep unaffected MODEL signatures, Sampling v6 guards, old unsafe/v5 history and official Workflow bytes unchanged. 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.
## 3.9.7 — external-input regeneration cache (2026-10-15)
- Fix automatic Full Run variations when Chunks or Variation Nonce come from external nodes. Unresolved linked inputs no longer let Core skip an explicitly requested regeneration. Earlier accepted chunks, existing Takes, public inputs and official workflows are preserved. Update the custom nodes, restart ComfyUI and refresh the browser; no workflow changes or history deletion are required.
- CPU: 1,948 passed / 2 skipped / 2 known Inspector expected failures; zero unexpected failures and zero CUDA initialization attempts. Targeted Core-API GPU queues verified external Chunks=2 / Nonce=0, repeated identical automatic variations 1 then 2, and 1 reused / 1 generated chunk per variation. Browser save/reload and Run-button submissions were checked separately. GPU scope: PyTorch attention, Spectrum OFF, 512x512, 8-step Turbo, two 5-second chunks. This is a targeted cache repair, not general image-quality or all-accelerator acceptance.
## 3.9.6 — regeneration intent and automatic variation cache (2026-10-15)
- Preserve explicit Regenerate From and Variation Nonce across V3.8/V3.9 workflow restoration and Queue submission. Persist only the selected intent as workflow metadata; keep the legacy unowned Chunk 1 / nonce 0 cleanup and one-shot restart behavior.
- Execute Full Run explicit regeneration with automatic nonce 0 through Run Storage on every Queue. Ordinary Full Run and fixed positive nonces retain Core caching. Sampling v6 safety, accepted Takes, public inputs and official workflow files are 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.
## 3.9.5 — Issue #31 Attention and Spectrum reuse signatures (2026-10-15)
- Observe the loaded Core PyTorch and Comfy Kitchen INT8 attention overrides with a finite local descriptor, including the verified wrapper self-reference, implementations, defaults, closure, both container dispatch paths, numerical helpers and provider state.
- Keep Sampling v6, runtime signature v1 and Graph v4 unchanged. Overrides outside this adapter retain their existing observation policy; unrelated incomplete MODEL/CFG observations still disable reuse.
- Add a dedicated Spectrum binding identity for explicit audited provider implementations, including upstream Spectrum 0.2.29. Include every effective immutable config value and verified loaded functions/wrappers; exclude reviewed execution-local history and counters. Unknown provider sources, extra runtime state or altered callables remain non-reusable.
- Preserve original unsafe Takes and v5 history. Only new correctly identified runs can gain reusable prefixes. Public nodes, workflow bytes and storage schemas are unchanged. See the versioned validation record for CPU/GPU coverage and remaining limits.
## 3.9.4 — Registry packaging repair (2026-10-14)