Skill optimization pass: routing, fail-closed gates, colour pipeline, memory integrity and canonical backport

.github/CONTRIBUTING.md

@@ -34,6 +34,24 @@ python src/release_manager.py selftest
 
 PR CI 會在三種作業系統與兩個 Python 版本驗證，另有可重現建置檢查。測試通過只代表已覆蓋的條件；未測的環境、資安或媒體品質應如實列出。
 
+## 維護者：把合併的公開修改同步回私有 canonical
+
+公開版的 derived 檔案（`src/`、`codex-skill/` 等由 canonical 投影出來的檔案）在 PR 合併後，要先回寫到維護者本機的私有 canonical，下一次 `sync_canonical.py` 才不會把修改蓋掉。用 `scripts/canonical_backport.py`：
+
+```text
+python scripts/canonical_backport.py --canonical <私有 canonical 路徑> --base <上次同步的 commit> --head <要同步的 commit>
+python scripts/canonical_backport.py --canonical <私有 canonical 路徑> --base <...> --head <...> --apply
+python scripts/sync_canonical.py --canonical <私有 canonical 路徑>
+git status
+```
+
+- 第一行是 dry run：逐檔列出 `FAST_FORWARD`／`MERGED`／`CREATE`／`DELETE`／`ALREADY_APPLIED`／`SKIPPED`／`CONFLICT`。`SKIPPED` 是不由 canonical 投影的檔案（repo-native、public-owned、generated、projection），不需要回寫；`tools/code-cleanup-helper/` 的修改要另外加 `--cleanup-helper <路徑>`；`docs/AUTOPILOT_ARCHITECTURE_V6.md` 來自私人工作區（`sync_canonical.py` 從公開 kit 資料夾的上一層讀它），要加 `--workspace <那個資料夾>` 才會回寫。
+- 有 `CONFLICT` 或 `ERROR` 時 `--apply` 會拒絕；處理衝突後重跑，或明確加 `--partial` 只寫其餘檔案。`--apply` 前會備份到 `--backup-dir`（預設是公開 repo 內、已被 git 忽略的 `.video-autopilot/backport/<時間>/`），寫入失敗會整批還原，並保留原檔的 BOM 與換行。
+- 公開 `SKILL.md` 由 `scripts/public_skill_skeleton.py` 產生，不會自動回寫；`--apply` 會在備份資料夾輸出 `SKILL.public-delta.diff` 與 `SKILL.public-head.md`，請手動把需要的段落併進私有 `SKILL.md`。
+- 工具逐個 first-parent commit 重播，公開版專用的改寫規則（路徑、隱私 sanitizer）不會被寫回 canonical。PR 請用 merge commit（不要 squash），或直接對 PR 分支執行，才保留逐 commit 的重播資訊。
+- 回寫並重新同步後，`git status` 應只剩 `sync-receipt.json` 變動；若還有其他差異，代表回寫不完整，先停下來檢查。
+- 已套用過再跑一次會回報全部 `ALREADY_APPLIED`。`python scripts/canonical_backport.py --self-test` 驗證工具本身。
+
 ## 隱私與安全
 
 不要提交 `config.py`、`profiles/`、`projects/`、頻道後台資料、客戶資訊、未公開影片、API key／token、session 紀錄或帶存取權限的遠端審片連結。範例使用虛構資料，證據先去識別化。

CHANGELOG.md

@@ -1,3 +1,116 @@
+## Unreleased
+
+### Skill routing and sync-back
+
+- The public `SKILL.md` now routes every one of its 56 references by task
+  (large references by section), starts each task from creator memory and a
+  context packet, and states a definition of done. The trigger description
+  covers Chinese requests (剪片、上字幕、調色、Shorts/Reels、封面、審片、發布).
+  The agent card no longer offers a remote review link by default.
+- New `scripts/canonical_backport.py` writes merged public edits back into the
+  private canonical before the next sync: per-commit replay, three-way merge
+  that keeps BOM and line endings, public-only rewrites never written back,
+  all-or-nothing `--apply` with backups, and idempotent re-runs.
+- New `scripts/hao_autopilot.py` command hub lists and runs the kit's entry
+  points.
+- Public docs name the reviewing creator by role (創作者 / the creator);
+  brand names, schema ids and the repository owner are unchanged.
+
+### Review integrity
+
+- Saved reviews carry an HMAC over the exact render's sha256; `finalize`
+  re-hashes the render, requires a submitted stamped review and archives stale
+  reviews on rebuild. Quality-95 certifies only from a verified review.
+- Missing evidence is an explicit REVIEW gap, never a silent full score; NaN,
+  infinite, boolean or text ratings never score.
+- `publish_contract` and `publish_hub` treat only `CERTIFIED_95` as
+  certified or ready.
+
+### Gates fail closed
+
+- Gate inputs: typed interview sources (`url:`, `file:`, `sha256:`),
+  placeholder detection shared by every gate, a dated compliance stamp with a
+  named reviewer, finite positive shot durations, proof images matched by real
+  path or content hash, a hash-frozen parity contract and split-battle Shorts
+  ids. `gate_core` raises when a report says not-ok or has no `fails` list, and
+  a self-test with zero checks is red.
+- Media QA: WCAG photosensitive-flash detection, long black runs, letterboxing
+  anywhere in the video, true peak <= -1.0 dBTP for Shorts, missing streams,
+  `-inf` loudness, special-character filenames and missing caption tools are
+  failures or UNVERIFIED, and anything not PASS blocks delivery.
+- Workflow controller: `workflow resume <run> --reaudit` recovers expired or
+  spent audit receipts; the state lock is owner-token based and survives dead
+  owners; the sealed plan is re-verified at audit, apply and verify; real words
+  containing "todo" are no longer rejected.
+- `script_gate.load_vocab()` is back, so creators can load their own audited
+  audience vocabulary; the restored Shorts gate API (`DEFAULT_RULES`,
+  `merge_rules`, `rules=`) runs example 04 again.
+
+### Memory and learning
+
+- Outcome baselines are per platform, format and window, count distinct
+  videos, and exclude early, late, undated, duplicate or non-finite captures
+  with a data-quality report.
+- The taste model rebuilds ratings from the vote log and needs at least five
+  distinct comparisons plus a 95% Wilson margin before declaring a preference.
+- Channel-state and taste writes are atomic and run under a lock, asset-usage
+  compaction is crash-safe, and interview scheduling and the publishing hub
+  share `data/channel_state.json`.
+- Learned rules move to `data/knowledge_state.json` (`VIDEO_AUTOPILOT_KNOWLEDGE_STATE`
+  overrides it), seeded once from the shipped `knowledge/runtime/state.json`, so
+  an update can no longer overwrite learning; support counts distinct sources,
+  so a retried review cannot promote a rule.
+
+### Colour and effects
+
+- Every RGB-domain filter (LUTs, curves, colour balance, channel mixer) runs
+  inside explicit BT.709 conversions and outputs are tagged BT.709. An identity
+  LUT now reproduces SMPTE bars at about 70 dB instead of 29.5 dB, so reds no
+  longer drift from (191, 0, 0) to (205, 17, 0).
+- One HDR-to-SDR normaliser (npl=100) is shared by the Shorts, grading and
+  tracking paths.
+- Filter renderers use lossless intermediates instead of `mp4v`; scanlines
+  darken, halftone coverage follows tone, whip and chromatic shifts replicate
+  edges instead of wrapping, and bloom no longer darkens the background.
+- Ken Burns helpers parse again, stay centred, move for the full clip and keep
+  the target aspect ratio.
+
+### Captions and typesetting
+
+- Shorts captions measure real libass widths and break only at rule-safe points:
+  never inside a Latin word or number, never with closing punctuation at a line
+  start. Lines scale down to 70% at most and raise instead of overflowing, and
+  every caption, chip and overlay stays inside one text-safe rectangle.
+- Text on accent colours picks ink or white by WCAG contrast (at least 5.13:1
+  across all themes); a missing CJK font raises instead of drawing tofu.
+- The thumbnail gate measures linearised WCAG contrast and keeps the headline
+  clear of the duration badge.
+- Shorts renders keep HDR at full precision until tone mapping, convert ungraded
+  clips with an explicit BT.709 matrix, blend cards and motion cues in RGB and
+  tag every encode BT.709.
+
+### Updater
+
+- An interrupted update (Ctrl+C, a killed process) rolls back automatically;
+  the next `check`, `auto` or apply reports
+  `INTERRUPTED_TRANSACTION_RECOVERED`, or `INTERRUPTED_TRANSACTION` when manual
+  review is safer. Apply and rollback hold an install-root lock (`LOCKED`).
+- Windows restarts after an update through a child process; archive order is
+  identical on every OS; errors keep their message unless it contains an
+  absolute path.
+
+### Docs, rules and health
+
+- Reconciled contradictions with the code: colour order and LUT strength,
+  review serving (local by default, phone review only by explicit request),
+  tutorial pacing (bounded by `pace_gate`), master loudness (-14 LUFS,
+  <= -1 dBTP), Signal Grid as a legacy family, the voice-profile rule, the
+  thumbnail test plan and the shipped design DNA seed count.
+- `system_health --quick` now also runs the gate, memory, backport, command hub
+  and media-QA self-tests plus examples 02/04/05/06, and fails on annotations
+  that would crash Python 3.9 at import time.
+- `.github/CONTRIBUTING.md` documents the maintainer backport workflow.
+
 ## v0.23.0 (2026-08-27)
 
 ### Architecture cleanup and public release closure

README.en.md

@@ -1,7 +1,7 @@
 # 🎬 video-autopilot-kit
 
 > **v0.23.0 / architecture 7.0**: long-form, Shorts and Reels share a self-authored composition runtime, an Imagegen-first material gate,
-> evidence-gated cinematic craft and one 38-preset filter library. Camera, edit, sound, color, VFX and transitions fall back to a clean cut when shot evidence is missing. The 33-reference design DNA, tracking, Quality-95 review, publishing packages, outcome learning and safe updates for long-form,
+> evidence-gated cinematic craft and one 38-preset filter library. Camera, edit, sound, color, VFX and transitions fall back to a clean cut when shot evidence is missing. The 16-seed public design DNA, tracking, Quality-95 review, publishing packages, outcome learning and safe updates for long-form,
 > Shorts and Reels. Run `python src/system_health.py --quick` to validate a clean install. Personal
 > media and analytics never enter the release.
 
@@ -69,20 +69,22 @@ misleading) or stay too generic to have real methodology. This kit gives you the
 (a battle-tested structure); `SETUP.md` **asks you questions** one section at a time, and
 your answers fill it in — so it actually becomes **your** system.
 
-## 🆕 New in v0.12.0 — borrowed numbers, evicted
+## Borrowed numbers, evicted (a design rule since v0.12)
 
-This release **removes no capability. It removes borrowed certainty.** Four unrelated parts of
+v0.12 **removed no capability. It removed borrowed certainty.** Four unrelated parts of
 the kit were making the same mistake: a number nobody measured, wearing an authority label, is
 worse than no number at all — because you trust it.
 
 - **The Shorts duration band is platform-aware.** Its dead zone was measured on **YouTube**
   Shorts; applied to IG/FB it blocks cuts that perform perfectly well. Pick a band with
   `spec["platform"]` (`rules=` still wins per key); an unknown platform name is a **blocking
   failure**, never a silent fallback.
-- **The script gate's four audience word-lists now ship empty.** Jargon grading can only be
-  audited out of **your own** transcripts — copy someone else's whitelist and you are checking
-  your script against their audience. An empty vocab doesn't block you (one warning, that's it);
-  `load_vocab()` loads yours.
+- **The script gate's four audience word-lists are only a starter set.** The kit ships a generic
+  starter vocabulary (`PUBLIC_FIXTURE`); jargon grading can only be audited out of **your own**
+  transcripts — copy someone else's whitelist and you are checking your script against their
+  audience. `script_gate.load_vocab("your audience_vocab.json")` replaces all four lists (skeleton:
+  [`templates/audience_vocab.example.json`](templates/audience_vocab.example.json)); an empty vocab
+  is not scanned and returns a single `lang.no_vocab` warning.
 - **A compliance layer + a "no link, no number" rule for the algorithm line** —
   [`knowledge/ai-content-compliance.md`](knowledge/ai-content-compliance.md) (R26-R38 + a 10-item
   pre-publish checklist) with 53 graded citations. Thresholds with no findable official source are
@@ -194,6 +196,11 @@ python install_or_upgrade.py --install-root <your-folder> --apply --install-skil
   managed file has local edits.
 - Every replacement is backed up under `.video-autopilot/backups/<transaction>/`; use
   `python src/release_manager.py rollback` to restore it.
+- An interrupted update (Ctrl+C, a closed terminal) rolls back automatically: the next `check`,
+  `auto` or `--apply` restores the previous version and reports `INTERRUPTED_TRANSACTION_RECOVERED`
+  (nothing new is applied that run; run it again). If recovery is not provably safe it reports
+  `INTERRUPTED_TRANSACTION` with the transaction id for manual review and `rollback`. Only one
+  update runs at a time; a second one gets `LOCKED`.
 - Major or compatibility-window-breaking releases always require confirmation.
 
 See the full contract in

README.md

@@ -3,7 +3,7 @@
 > Current version: v0.23.0
 >
 > **v0.23.0／架構 7.0**：長片／Shorts／Reels 共用自研 composition runtime、Imagegen-first 材質閘門、證據式電影工藝、統一濾鏡庫與唯一發布中樞；
-> 38 組調色／動態／轉場／主體濾鏡、33 圖設計 DNA、Tracking、Quality-95 審片、發佈包、成效學習與安全自動更新都在同一套公開版。沒有鏡頭證據時仍會退回乾淨切鏡。
+> 38 組調色／動態／轉場／主體濾鏡、16 組公開設計 DNA 種子、Tracking、Quality-95 審片、發佈包、成效學習與安全自動更新都在同一套公開版。沒有鏡頭證據時仍會退回乾淨切鏡。
 > 執行 `python src/system_health.py --quick` 可驗證乾淨安裝；個人媒體與成效資料不會進 release。
 
 > 一套**框架式**的 YouTube / 短影音自動化工具 + 方法論模板。
@@ -62,15 +62,16 @@ python src/filter_runtime.py gallery a.mp4 review/filter-library --source-b b.mp
 要嘛太通用沒有方法論。這個 kit 給你**骨架**（經實戰的結構），
 `SETUP.md` 一區一區**問你問題**，用你的答案填滿它 —— 這樣它才真的是**你的**系統。
 
-## 🆕 v0.12.0 新增 — 把「借來的數字」清出去
+## 把「借來的數字」清出去（v0.12 起的設計原則）
 
-這一版**沒有拿掉任何功能，拿掉的是借來的把握**。四個地方犯的是同一種錯：
+v0.12 **沒有拿掉任何功能，拿掉的是借來的把握**。四個地方犯的是同一種錯：
 一個沒人量過的數字，掛上權威標籤，比沒有數字更糟 —— 因為你會信它。
 
 - **Shorts 片長帶改平台感知** —— 死區是在 **YT Shorts** 上量出來的，套到 IG/FB 會擋掉正常的剪法。
   改用 `spec["platform"]` 選帶（`rules=` 仍逐鍵優先）；平台名打錯是**擋下的失敗**，不是靜默 fallback
-- **腳本 gate 的四層詞表改成出貨即空** —— 行話分級只能從**你自己的逐字稿**審計出來；
-  照抄別人的白名單 = 用別人的觀眾檢查你的稿。空表不擋你（只回一條 warn），`load_vocab()` 載你自己的
+- **腳本 gate 的四層詞表只是起手式** —— 出貨的是通用 starter 詞表（`PUBLIC_FIXTURE`）；行話分級只能從**你自己的逐字稿**審計出來，
+  照抄別人的白名單 = 用別人的觀眾檢查你的稿。`script_gate.load_vocab("你的 audience_vocab.json")` 整組換成你的
+  （骨架：[`templates/audience_vocab.example.json`](templates/audience_vocab.example.json)）；載入空表時不掃詞，只回一條 `lang.no_vocab` warn
 - **演算法線補上合規層＋「沒出處就不引用」** —— [`knowledge/ai-content-compliance.md`](knowledge/ai-content-compliance.md)（R26-R38 ＋ 發布前 10 項 checklist）
   ＋ 53 條分級法源；查無官方出處的門檻數字就地標記，不再與有出處的並排
 - **新工具 [`src/teardown.py`](src/teardown.py)** —— 一個指令把競品直式短片拆成可比較的數字（刀速／刀距分布／換句速率／換句÷剪點／LUFS）；
@@ -169,6 +170,9 @@ python install_or_upgrade.py --install-root <你的資料夾> --apply --install-
 - 未知自訂檔永不刪；已修改的官方管理檔在自動模式會停在 `CONFIRM_REQUIRED`。
 - 每次覆蓋前建立 `.video-autopilot/backups/<transaction>/`；需要時執行
   `python src/release_manager.py rollback`。
+- 更新中斷（Ctrl+C、程式被關掉）會自動回滾：下一次 `check`／`auto`／`--apply` 先回到舊版並回報
+  `INTERRUPTED_TRANSACTION_RECOVERED`（這次不套新版，重跑一次即可）；無法安全判斷時回報
+  `INTERRUPTED_TRANSACTION` 與交易編號，請人工檢查後 `rollback`。同時只能有一個更新在跑，另一個會得到 `LOCKED`。
 - 重大／不相容版本不會靜默升級，必須由使用者確認。
 
 完整契約見 [`codex-skill/video-autopilot/references/open-source-release-and-upgrade.md`](codex-skill/video-autopilot/references/open-source-release-and-upgrade.md)。

SETUP.en.md

@@ -104,9 +104,10 @@ The AI asks one question at a time and fills the files for you — **you just an
 > into things a **machine** can block? Use the cumulative version,
 > `templates/style_profile.template.md`. Its §5 produces `audience_vocab.json`
 > (skeleton: `templates/audience_vocab.example.json`), which
-> `src/longform_maker/script_gate.py` reads to gate a script **before you record it**.
-> **Those four word-lists ship empty on purpose** — they can only be audited out of your own
-> transcripts, and copying someone else's means checking your script against *their* audience.
+> `src/longform_maker/script_gate.py` uses, after `script_gate.load_vocab("your audience_vocab.json")`,
+> to gate a script **before you record it**. **The built-in lists are only a generic starter set** —
+> real grading can only be audited out of your own transcripts, and copying someone else's means
+> checking your script against *their* audience.
 > Method → `knowledge/script-retention-craft.md`.
 
 ## 4️⃣ Production → generates `config.py`　★required
@@ -161,8 +162,11 @@ prints a `WARN 節目 profile 未填: …` line listing exactly which fields are
 > this verbatim, so an incomplete list means an incomplete release**).
 >
 > ⚠️ **Only a human can stamp compliance**: `plan` writes the compliance field as "pending review"
-> and the gate blocks on it; you pass `--compliance-ok` yourself, after you've walked your
-> platform's AI-content policy checklist. Methodology →
+> and the gate blocks on it; you pass `--compliance-ok --reviewer "Your Name"` yourself, after
+> you've walked your platform's AI-content policy checklist (the stamp carries the date and the
+> reviewer; without `--reviewer` it uses `HOST_NAME` from `profiles/show.md`, and blocks if both
+> are empty). `CLUSTER` must hold a real topic keyword too — a template placeholder fails P5.
+> Methodology →
 > [`knowledge/interview-show-playbook.md`](knowledge/interview-show-playbook.md).
 
 ## 8️⃣ Shorts rule calibration → override the `shorts_gate` thresholds　⭕optional (**only if you cut vertical Shorts**)