Update vendored H3 Continuum to v3.9.11

This commit is contained in:
2026-10-20 20:21:06 +00:00
parent c8568633a4
commit 494670ded0
26 changed files with 625 additions and 77 deletions
@@ -1,5 +1,66 @@
# Issue #31 — Attention and Spectrum reuse signatures (3.9.5)
## 3.9.11 — stable supplemental loaded-code audit
Repeated imports of unchanged audited Spectrum source can produce different
marshal bytes for equivalent normalized code objects, observed with Python
3.13.12. This can mark a new Run read-only even though no source or settings
changed. It is a signature observation failure; the reporter's exact environment
has not been reproduced.
The original audit remains the first path. Only differences in function
`code_sha256` fields can trigger a supplemental audit. Fixed
`spectrum_code_audit_v1_py312.json` and `spectrum_code_audit_v1_py313.json`
profiles check all loaded provider module descriptors: typed code fields and
nested constants, defaults, keyword defaults, closures, class members and
provider references. Float/complex bits and constant types are preserved.
Source hashes, exact Python file sets, module provenance and runtime-state
checks remain required. Unknown constants or missing/invalid profiles reject
reuse; independent Attention failures remain failures.
The profiles are bound to the original metadata digest and complete audited
source set. They are generated offline with `tools/build_spectrum_code_audit.py`,
using explicit audited provider roots; opposite provider/module import orders
must produce identical bytes. The runtime never builds a profile, executes
provider functions, imports Spectrum or learns an approval from current code.
All four legacy provider metadata files and the returned persistent identity
remain unchanged. Normal pre-repair v6 Takes retain compatibility. Stored unsafe
Takes and v5 history stay readable and read-only; the repair never edits their
flags or manifests. Use a new Run when checking a previously rejected setup.
Sampling v6, Graph v4, Run Storage, public inputs and shipped workflows are
unchanged. Targeted browser/GPU Review continuation passed with the audited
local Spectrum provider and Kitchen Attention: Chunk 1 retained, Chunk 2
generated, one reused / one new, both AV outputs fully decoded. The supplemental
path is proven by CPU reproduction; its use during GPU execution is not traced.
See RELEASE_NOTES_V3911.md for environment, limits and the separate third-party
MP4 menu-thumbnail workaround.
CPU/JS validation: **2,139 passed / 2 skipped / 2 known Inspector expected
failures**, zero unexpected failures or CUDA initialization attempts. This
includes the existing 90-case Queue JS fixture. An additional 53 real-Core
observer/storage cases matched the production UI's read-only predicate: 30
reusable and 23 intentionally rejected. These use tiny CPU fixtures and actual
Core function references; no Attention kernel or H3 inference runs. Four normal
v6 Takes saved by pre-repair code kept identical full contracts and reused
Chunk 1 without changing their source files. Existing tests also preserve
unsafe/v5 history and reject reuse through the storage entry points.
Python 3.12 component coverage passed 155 tests with one skip; 21 tests tied to
the older Core's separate Attention limitation were excluded. The storage matrix
still checks that environment's eight Kitchen cases remain non-reusable for
its unobservable helper. This repair does not resolve that independent failure.
### 日本語 — 3.9.11の補助監査
同じ監査済みSpectrumを再読込した際、同等のコードでもmarshalの保存表現が変わり、
再利用不可になる条件へ対応します。旧監査を優先し、関数のコードハッシュだけが
異なる場合に限り、固定プロファイルでコードと読込済み記述全体を再検査します。
未知のコード・既定値やclosureの変更・追加状態の拒否は維持します。
旧署名を変更せず、正常な既存v6 Takeの互換性を保ちます。過去のunsafe/v5 Takeを
自動的に再利用可能にはしません。対象を限定したSpectrum+Kitchenの実ブラウザー・GPU検証で、
Review操作とChunk 1保持/Chunk 2生成を確認しました。詳細と制約はRELEASE_NOTES_V3911.mdを参照。
## 3.9.10 — Spectrum audit portability
The repair treats source CRLF and LF line endings as equivalent
+98
View File
@@ -0,0 +1,98 @@
# v3.9.11 — Spectrum signature stability / Review continuation
## What changed
Fix an additional confirmed cause of false Sampling v6 read-only. Repeated
imports of unchanged audited Spectrum source can produce different marshal
bytes for equivalent normalized code objects, reproduced on Python 3.13.12.
The legacy audit remains first. Only differences confined to function code
digests can invoke fixed Python 3.12/3.13 supplemental profiles; all loaded
descriptors, source hashes, provenance, defaults, closures and runtime state
must still pass. Unknown or modified code remains non-reusable.
The original four provider metadata files and persisted Sampling v6 identity
are unchanged. Four genuine pre-repair safe v6 Take fixtures retained identical
full contracts and reused their saved Chunk 1. Old unsafe/v5 Takes stay readable
and read-only; no flags or manifests are rewritten. Public node IDs, inputs,
widget order, Graph v4, Run Storage and official workflow JSON/ZIP bytes remain
unchanged. No provider import, profile generation or identity learning occurs
in the runtime audit.
## Validation
- Full CPU/JavaScript gate: **2,139 passed / 2 skipped / 2 known Inspector
expected failures / zero unexpected failures**, zero CUDA initialization.
Existing 90-case Queue JS coverage is included, not an additional total.
- Targeted Python 3.13: 212 passed / 1 skipped. Python 3.12.10 component gate:
155 passed / 1 skipped / 21 deselected. Those 21 depend on an independent
older-Core Attention observation limitation and are not accepted here.
- Real-Core CPU observer/storage fixtures matched 53 expected decisions
(30 reusable, 23 intentionally refused) and the production JS predicate.
These fixtures execute no Attention kernels or H3 inference. Old unsafe/v5
refusal and changed/unknown-code rejection remain tested.
- Both supplemental profiles regenerated byte-for-byte with reversed provider
and module import order. Compilation and offline twelve-node registration
passed. Legacy metadata and official workflow bytes remained unchanged.
- **Browser + GPU:** Review buttons enabled with Spectrum + Comfy Kitchen
Attention. The browser's Continue From Here action retained **Chunk 1** and
generated **Chunk 2**, **one reused / one new**. Both new Takes were reusable,
parent links were correct, and the saved Chunk 1 raw/Take/manifest stayed
unchanged. Two Queues completed without execution errors or OOM; both MP4s
fully decoded video and audio (5 s / 120 frames; 10 s / 240 frames).
GPU scope: Windows, Python 3.13.12, ComfyUI Core 0.42.1, PyTorch 2.14.1+cu130,
Comfy Kitchen 0.2.39, RTX 5060 Ti 16 GB, audited installed 44-file local Spectrum
provider ON, Kitchen launch option and ModelAttentionBackend selected, Sage OFF.
T2VA 512×512, 24 fps, two 5-second chunks, 8-step Turbo, Euler/simple, first pass
only. Audited KJ Preview Override was enabled with optional VAEs disconnected
and Tiny VAE none. This tests the repaired product payload before release
branding; it is not a fresh GPU run of the published tag.
No reporter-exact reproduction, all-provider GPU matrix, Second Pass acceptance,
subjective visual/audio quality, benchmark or CUDA-kernel return tracing is
claimed. The supplemental path was demonstrated by CPU reproduction; whether it
was invoked in the live GPU run was not traced. Older Core 0.39.2 with Kitchen
0.2.37 has a separate unobservable prequantize_int8_attention helper; that setup
remains non-reusable and is not fixed by this release.
## Update and check
Update the existing copy through ComfyUI Manager, or replace the custom-node
package in `ComfyUI/custom_nodes/ComfyUI-H3-Continuum`. For a Git installation,
update that same folder from main. Avoid duplicate installations. **Fully restart
ComfyUI and hard-refresh the browser with Ctrl+F5.** Keep existing Takes,
workflows and model files.
Use a **new, unique Run Name**, enable Spectrum and Review Each Chunk, and
generate Chunk 1. Select its new Take in Render History; check Use This Take /
Continue From Here, then continue to Chunk 2. Existing read-only Takes are not
automatically made reusable. No workflow rewiring or Take deletion is required.
## Separate MP4 menu-thumbnail workaround
Custom Scripts' Show Image On Menu can request an MP4 through Core's image
preview endpoint, causing an HTTP error despite successful generation and valid
video files. If encountered, search Settings for **Show Image On Menu** and
turn it **OFF**. A browser load/save of an existing MP4 after this setting change
produced **zero new thumbnail HTTP errors**. The test performed no new inference.
This is a separately verified setting workaround, not a Continuum/Core/Custom
Scripts code change; this update does not alter the setting automatically.
## 日本語
同じ監査済みSpectrumを再読込した際、コードが同等でもmarshalの保存表現が変わり、
新規Takeが誤ってv6 read-onlyになる条件を修正しました。旧監査を優先し、コードハッシュ
だけが異なる場合に限り固定プロファイルで全記述を照合します。未知・改変コードの
再利用拒否、元のSampling v6署名、正常な既存Takeの互換性を維持します。
過去のunsafe/v5 Takeを自動的に再利用可能にせず、既存Take・公式ワークフローを変更しません。
CPU/JSは2,139成功・2スキップ・既知Inspector期待失敗2件、予期しない失敗0。
Spectrum+Kitchen Attentionの対象を限定したブラウザー/GPU検証ではReviewボタンが
有効となり、Chunk 1保持・Chunk 2生成(再利用1/新規1)と両出力の映像・音声の全デコードを確認。
検証は上記構成の修正版payloadに対するもので、公開タグのGPU再実行、全構成、画質・速度、
報告者環境の完全再現を保証するものではありません。旧Core/Kitchenの独立した制約も残ります。
Managerまたは既存custom-nodeフォルダーを更新し、ComfyUIを完全に再起動、Ctrl+F5で更新。
一意な新しいRun NameでSpectrumとReview Each Chunkを有効にしてChunk 1→2を確認してください。
MP4メニューサムネイルの別拡張由来HTTPエラーは、SettingsのShow Image On MenuをOFFに
する回避方法を確認済みです。コード修正や設定の自動変更は含みません。