@chrono-meta/fh-gate 3.2.0 → 3.5.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrono-meta/fh-gate",
-  "version": "3.2.0",
+  "version": "3.5.0",
   "description": "FH runtime adapters — run FH governance, skills, and agents via Claude or Codex with machine-parseable gates.",
   "homepage": "https://github.com/chrono-meta/forge-harness#readme",
   "bugs": {
@@ -28,6 +28,7 @@
     "fh-codex-doctor": "bin/fh-codex-doctor.js"
   },
   "scripts": {
+    "prepack": "find plugins -type d -name __pycache__ -prune -exec rm -rf {} + 2>/dev/null; true",
     "prepare": "chmod +x bin/fh-gate.js bin/fh-run.js bin/fh-goal.js bin/fh-codex-doctor.js scripts/fh-gate.sh scripts/fh-run.sh scripts/fh-goal.sh",
     "postinstall": "node scripts/postinstall_notice.js",
     "test": "bash scripts/selfcheck.sh",
@@ -75,6 +76,9 @@
     "scripts/version_lockstep_check.sh",
     "scripts/test_selfcheck_state_lanes.sh",
     "scripts/test_lane_runner_lanes.sh",
+    "scripts/usage_report.py",
+    "scripts/test_usage_ledger_lanes.sh",
+    "scripts/test_fh_run_resolver_lanes.sh",
     "scripts/test_version_lockstep_lanes.sh",
     "scripts/package_coverage_check.sh",
     "scripts/files_manifest_shipping_check.sh",
@@ -227,6 +231,7 @@
     "plugins/fh-meta/skills",
     "plugins/fh-meta/agents",
     "plugins/fh-commons/.claude-plugin/plugin.json",
+    "plugins/fh-commons/README.md",
     "plugins/fh-commons/skills",
     "plugins/fh-commons/agents",
     "knowledge/shared/harness-core",
@@ -322,6 +327,9 @@
     "scripts/test_utterance_intake_lanes.sh",
     "scripts/worktree_reclaim.sh",
     "scripts/test_worktree_reclaim_lanes.sh",
+    "plugins/fh-preprep/.claude-plugin/plugin.json",
+    "plugins/fh-preprep/README.md",
+    "plugins/fh-preprep/skills",
     "plugins/fh-qp/.claude-plugin/plugin.json",
     "plugins/fh-qp/README.md",
     "plugins/fh-qp/qp_profile.example.yaml",

package/plugins/fh-commons/.claude-plugin/plugin.json

@@ -1,10 +1,10 @@
 {
   "name": "fh-commons",
-  "version": "3.2.0",
+  "version": "3.5.0",
   "engines": {
     "claudeCode": ">=1.0.0"
   },
-  "description": "Cross-project utility skill bundle — 6 skills + 1 agent. Domain-agnostic and portable: convergence-loop, deliberation, mcp-circuit-breaker, token-budget-gate, ko-tech-writer, preprep + quench-challenger agent.",
+  "description": "Cross-project utility skill bundle — 5 skills + 1 agent. Domain-agnostic and portable: convergence-loop, deliberation, mcp-circuit-breaker, token-budget-gate, ko-tech-writer, preprep + quench-challenger agent.",
   "author": {
     "name": "chrono-meta",
     "email": "chrono-meta@users.noreply.github.com"

package/plugins/fh-commons/README.md

@@ -0,0 +1,38 @@
+# fh-commons — Contention-Born Skill Habitat
+
+**A different layer of plugin from fh-meta.**
+
+fh-meta = meta-engineering skills for operating, diagnosing, and improving the harness itself  
+fh-commons = general-purpose utility skills born from the contention layer, transplantable to any project
+
+## Placement Criteria
+
+When the `contention-layer` skill harvests contentions and generates new skill candidates:
+
+| Determination | Destination |
+|---|---|
+| Harness engineering in nature | `fh-meta` |
+| Project-general · domain-agnostic | **`fh-commons`** (this plugin) |
+| Domain/team specific | field harvest (decided by the field team) |
+
+## Skill List
+
+| Skill | Description | Contention Parent |
+|---|---|---|
+| `convergence-loop` | General-purpose gate reinforcement that replaces single-pass structures with a convergence loop of up to N rounds | harvest-loop (recurring single-pass-distrust pattern across hub gates) |
+| `deliberation` | Innovator → Devil-Advocate → Mediator 3-layer multi-perspective synthesis. Generates conditional verdicts without binary win/loss | Migrated from fh-meta (2026-05-23 — domain-agnostic general decision structure) |
+| `ko-tech-writer` | Korean technical-writing pipeline — register detection, venue calibration, translationese scan, honesty layering, number/claim gate, perceptual QA | Field-session harvest (2026-08-10 — conference tech-doc revision session, operator-driven shipping) |
+
+## The `origin` Field
+
+Commons skills include the following fields in their SKILL.md frontmatter:
+
+```yaml
+# Contention-born (extracted by contention-layer)
+origin: contention-layer
+contention-parents: [skill-A, skill-B]
+
+# Migrated from fh-meta (reclassified as general-purpose)
+origin: fh-meta
+migration: "YYYY-MM-DD — reason for migration"
+```

package/plugins/fh-meta/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "fh-meta",
-  "version": "3.2.0",
+  "version": "3.5.0",
   "engines": {
     "claudeCode": ">=1.0.0"
   },

package/plugins/fh-meta/CHANGELOG.md

@@ -1,5 +1,122 @@
 # forge-harness (fh-meta) Changelog
 
+### [3.5.0] — 2026-09-13 — preprep 를 자기 플러그인으로 승격 (`fh-preprep`)
+
+**BREAKING (gate):** 스킬 경로가 **이동한다**.
+`plugins/fh-commons/skills/preprep/` → `plugins/fh-preprep/skills/preprep/`
+🟥 **그 경로를 직접 부르던 스크립트·문서·설정은 깨진다.** 처방 = 경로 한 곳 치환:
+`sed -i '' 's|fh-commons/skills/preprep|fh-preprep/skills/preprep|g' <파일>`
+플러그인으로 설치해 쓰는 경우(스킬 이름으로 호출)는 영향이 없다 — 이름은 `preprep` 그대로다.
+
+**왜 옮겼나**: 스킬 하나가 레인 17개(내장 7 + 모듈 10)·픽스처군·오라클을 들면서
+**역할이 갈렸다**(라우팅 / 판정 / 오라클). 규모가 아니라 역할 분기가 근거다.
+
+- 새 `plugins/fh-preprep/` — 자기 `plugin.json` · README(종료코드 계약 명시) · `skills/preprep/`
+- marketplace 항목 3 → **4**. 버전 락스텝 8개 문자열 전부 3.5.0
+- `files[]` 에 `plugins/fh-preprep/{plugin.json,README.md,skills}` 등재 — 빼먹으면 조용한 출하 누락
+- 드리프트 앵커·레인 12개·`field_canon_preload.sh`·`selfcheck.sh` 의 경로 참조를 같이 옮겼다
+
+**승격 전에 닫은 것**(순서가 근거다 — 앵커 수리와 이동이 한 diff 에 섞이면 귀속이 안 갈린다):
+① #702 머지로 드리프트 앵커 초록 · ② 실물 14 표면 완주(rc=1 확정) 로 Done-When #1·#6 닫힘
+(#705 에서 「설정을 못 읽음」이 `rc=1`(=발견 있음)으로 나가던 선재 결함을 `rc=2` 로 수리).
+
+🟥 **승격이 닫지 않는 것**: preprep Done-When #5(죽은 선언 12개)는 여전히 미충족이고 다른 축이다.
+
+### [3.4.0] — 2026-09-12 — 탈상관의 «강한 축» 을 기계로 옮긴다 (외부 실측 기인)
+
+**BREAKING (gate):** 마커 두 자리가 조인다. 둘 다 **grace date 로 소급하지 않는다** — 그 날짜 이전
+파일명의 마커는 종전대로다.
+① `crossfamily: panel(<families>)` 는 이제 grounds 에 **`evidence=SHARED|INDEPENDENT|MIXED(...)`**
+   토큰을 요구한다(`EVIDENCE_TOKEN_GRACE_DATE=2026-09-12`). 처방 = 한 토큰 추가:
+   `crossfamily: panel(codex) — residency=CLEAN(files=3) · evidence=SHARED(same staged diff to both) · …`
+② `standpoint:` 의 `tier2`/`tier2b`/`tier3` grounds 검사가 **advisory → 차단**
+   (`STANDPOINT_GROUNDS_GRACE_DATE=2026-09-12`). 처방 = 돌린 **명령과 본 출력을 명명**하거나,
+   실행을 안 했으면 정직하게 `tier1b`(grounds 미검사)로 내린다.
+🟥 **소비자 install 도 같다** — 설치된 훅은 그 레포의 마커를 같은 기준으로 본다.
+
+**왜 지금인가 — 근거가 내부 관찰에서 외부 실측으로 바뀌었다.** `arXiv:2609.10969`
+(*Engineering Reliable Commit Gates for Agentic AI: Cost-Aware Verification Portfolios under
+Common-Mode Data Failures*, 2026-09-10, cs.SE)이 탈상관 두 축을 **고정 호출예산 2×2**로 분리해
+48 템플릿 · **2,880 시나리오**에서 쟀다: 「같은 증거를 읽는 교차-모델 투표」는 위험 제안의 **62.9 %**
+를 승인하고 「독립 출처」는 **22.9 %** — 출처 효과 **40.9 %p** 대 모델 다양성 **11.3 %p**, **3.6 배**.
+즉 FH 가 **하드 차단하던 축이 약한 쪽**이었고, 강한 쪽의 grounds 는 advisory 였다. 그 불균형을 뒤집는다.
+
+**`SHARED` 는 합법이다.** 결함은 «같은 diff 를 읽은 패널» 이 틀렸다는 것이 아니라 그 사실이 **기록에
+적힐 수 없었다**는 것 — 그런 런이 이 enum 의 가장 강한 값으로 남았다. 차단되는 것은: 어느 쪽인지 안
+적은 `panel(...)` · 닫힌 셋 밖의 값 · 형식 위반/중복 토큰 · `SHARED`/`MIXED` 의 공허한 본문
+(`INDEPENDENT` 는 자기서술적이라 본문 검사 없음 — 정직한 답을 과차단하면 override 를 훈련시킨다).
+
+🟥 **막는 것은 «기록의 형태» 이고 «진위» 가 아니다.** 저자가 명령을 명명했는지는 기록의 속성이지만,
+그 실행이 진짜였는지는 이 게이트가 보지 않는다 — §Mechanization Boundary 가 사람에게 남긴 의도된 잔여다.
+
+**함께 — 도구 쪽 정직성 (`auto-decorrelation` Step 4.6 신설).** 🟥 **이 스킬의 기본형이 곧 약한 팔이다**:
+Step 4 가 payload 하나를 만들어 팬아웃하므로 평범하게 돌린 패널은 구조적으로 62.9 % 팔이다. 그 사실을
+스킬 본문에 적고, 토큰 값을 «의도가 아니라 실제 디스패치 형태로» 정하게 했다. **독립 출처 모드는 없다** —
+비용 구조·residency 경계·«독립» 의 정의가 전부 미결이라 설계 문제로 이름만 남겼다
+(`tracks/_meta/fh_signal_2026-09-12_independent-source-axis.md`, CHAMBER-CANDIDATE).
+
+**frontier-digest — 목록 경로의 ID↔제목 짝짓기.** 429 폴백 경로에서 **위치로 짝지으면 안 된다**:
+2026-09-12 실측에 목록이 **ID 51개 · 제목 50개**를 내서 두 병렬 추출이 중간부터 밀렸고, 그 뒤 짝은
+**형식은 멀쩩한데 전부 틀렸다**(`2609.10550` 외 2건). 한 항목을 **한 단위로** 읽고, 그게 확인되기 전까지
+**항목별 `abs/{id}` 재해소가 의무**이며 제목 불일치는 **버린다**. 훅은 없다(파싱이 세션 안에서 일어난다).
+
+🟥 **이 게이트의 초판이 여섯 군데 틀렸고, 다른 계열이 그걸 찾았다 (자력 적발 0).**
+같은 diff 를 `agy`(gemini-3.8-flash-high)에 보냈고 여섯 건 전부 **실행 프로브로 재현한 뒤** 고쳤다.
+방향이 중요하다 — **넷은 과차단**이었다. 새 게이트가 advisory 가 아니라 차단이 된 순간, 과차단은
+경고 한 줄이 아니라 `--no-verify` 를 훈련시키는 결함이 된다:
+
+| # | 부류 | 초판 동작 | 수리 |
+|:-:|---|---|---|
+| ① | **과차단 S** | `evidence=MIXED(codex(diff), gemini(repo))` 의 본문이 안쪽 `)` 에서 잘려 «공허» 로 판정 → 정당한 기록 하드 차단 | 중첩 **한 겹** 허용(두 겹은 여전히 형식 위반) |
+| ② | **fail-open S** | `evidence=INDEPENDENT()` 와 `INDEPENDENT(<what>)` 가 **통과** — «길이 미검사» 를 «본문 없어도 됨» 으로 접었다 | 빈 본문·자리표시자는 **세 값 모두** 차단, 길이 바만 SHARED/MIXED |
+| ③ | **과차단 S** | `standpoint:` 키워드에 뒤따르는 공백이 박혀 `cargo check` · `python scripts/eval.py` · `ran: ./ci.sh` 가 전부 차단 | 경계를 비-알파벳으로 + 흔한 러너 추가 + `30/30`·`rc=`·`./path` 도 실행 증거로 |
+| ④ | **fail-open S** | `EVIDENCE_TOKEN_GRACE_DATE` 미주입 시 grace 가 **조용히 꺼져** pre-grace 마커를 전부 차단 | 상수를 mdate 검사 **앞에서** `:?` 로 평가 → 크게 죽는다 |
+| ⑤ | **fail-open A** | pre-grace 마커의 `evidence=PARTIAL(...)` 이 조용히 통과 | 부재는 면제, **오작성은 면제 아님**(residency 와 같은 약속) |
+| ⑦ | **과차단 A** | `wc -w` 가 `same-diff` 도 `동일diff` 도 1 낱말로 세어 정당한 **한국어** 기록을 차단 | 길이 바를 **문자 수**로 |
+
+⑥(가드 단축평가 — `[ -n "$mdate" ] && [ ... ${VAR:?} ]` 에서 mdate 가 비면 `:?` 가 **아예 안 돈다**)도
+같이 고쳤다. 🟥 **가드가 있다고 적어두고 안 도는 것은 장식이다**(`[[feedback_anchor_can_be_decorative]]`).
+
+**레인**: `test_marker_crossfamily_lanes.sh` **82 픽스처**(신규 `e1`–`e15`, 위 여섯 건이 각각 컨트롤과 짝) ·
+`test_marker_standpoint_lanes.sh` `N8`~`N8i`. 되돌림 프로브 두 종 — evidence 요구를 떼면
+**정확히 6개**(e1·e5·e6·e7·e8·e8c)만 적색, standpoint 차단을 떼면 **N8 하나만** 적색.
+🟥 두 레인 러너가 새 grace 상수를 주입하지 않으면 기존 픽스처가 통째로 오판하는 자리였다(standpoint 쪽은
+실제로 깨졌다) — 러너에 주입을 넣고, 훅에는 `:?` 가드를 박아 부분 추출이 **조용히 통과하지 않게** 했다.
+
+### [3.3.0] — 2026-09-12 — 짝지음(paired) 측정: 한 round-1 을 두 분기가 나눠 먹는다
+
+**① 새 자산 — `--r1-only` · `--reuse-r1 <dir>` (`finding_fleet.sh` · `finding_pipeline.sh`).**
+round-1 을 **한 번** 돌려 두고, 그 **같은** round-1 에서 round-2 를 두 번 갈라 돌린다
+(`--reuse-r1 <dir> --round2` / `--round2-blind`). 왜 필요했나: «round-2 에서 상대 계열의 목록을
+보여주는가» 를 A/B 하려고 두 팔을 독립 실행했더니 **round-1 기저부터 달랐다**(146 vs 136, 7.4 %) —
+즉 한 변수 비교가 아니었다. 이 플래그들은 그 교란을 설득이 아니라 **설계**로 없앤다.
+부수 효과로 더 싸다 — round-1 을 두 번 돌리지 않는다(멤버콜 −25 %).
+
+기계 보증(레인 `test_finding_pipeline_lanes.sh` L115~L124, 122 → **132**):
+두 분기의 `findings_r1.jsonl` 이 공유 소스와 **바이트 동일** · 파이프라인이 r1 **sha256** 을 두 분기
+로그에 찍어 사후 대조 가능 · **빈 r1 재사용 거부**(미측정을 0 으로 렌더하지 않는다) · 옛 round-2
+잔재 제거 · 소스==out 거부 · `--reuse-r1` 단독 거부 · `--r1-only` 은 검증을 안 돌리므로 **씨앗 선언을
+거부**한다(선언된 컨트롤이 돈 적 없이 «통과» 로 읽히는 것을 막는다).
+
+**② fail-open 수리 (`--r1-only`, cross-family 적발).** `--r1-only` 이 **빈 round-1 을 «사용 가능»
+으로 통과**시켰다. 기전: `grep -c . f || echo 0` 은 빈 파일에서 «`0` 을 찍고 rc=1» 이라 폴백이 줄을
+하나 더 붙여 `_n1="0\n0"` 이 되고, `[ "$_n1" -eq 0 ]` 이 **구문오류로 거짓**이 되어 `R1_UNUSABLE`
+분기를 건너뛴다. ⚠️ **새 경로라 종전 동작 변화는 없다** — 기존 팔(플래그 미사용)의 경로는 불변이고
+레인 122개가 그대로 통과한다. 질문은 «폴백을 붙였나» 가 아니라 **«실패 시 stdout 이 비는가»** 다.
+
+**③ 출하 누락 수리 — `plugins/fh-commons/README.md` 를 `files[]` 에 명시.** npm 11 은 중첩 README 를
+암묵 포함했는데 **npm 12 는 뺀다**(실측: `npx npm@latest pack --dry-run` 에서 0건). 발행 경로가
+npm@latest 이므로 그대로 두면 3.2.0 에서 이미 빠져 나갔을 파일이다.
+
+**자릿수 근거**: 새 capability **instance** 가 아니라 새 플래그 둘 + 레인 10 + 출하 파일 1 —
+§Version-Digit-Policy 의 minor(새 자산 · 새 게이트 레인)다. **`BREAKING (gate):` 없음** — 소비자의
+게이트 수용을 깨지 않는다(새 플래그는 opt-in, 기존 호출 경로 불변).
+
+**④ harness-pr-reviewer (개명 + Axis 2·3 의무 배선) — 이 릴리스가 싣는다.**
+- **RENAME** `hub-cc-pr-reviewer` → `harness-pr-reviewer` (운영자 결정 2026-09-12): 실제 쓰임이 «현장 하네스가 자기 PR 을 FH 리뷰 능력으로 검증하는 standalone 창구» 라서. 옛 이름 발화·참조는 그대로 라우팅(별칭), 리다이렉트 스텁 없음(phantom-quench 선례).
+- Step 3.5 «Axis 2·3 의무 디스패치 레인» — 트리거 3항(8-matrix ❌ · load-bearing 경로 grep · 머지 권고 요청)을 PR 에서 계산, Axis 2 = auto-decorrelation 재사용 + `crossfamily:` 닫힌 enum, Axis 3 = phantom-quench(N/A 는 `--name-only` 로), 판정형 질문 reps≥3. Self-Catch 는 «판정이 아니라 cue». 플로어 sim: 현행 0/3 → 수정 3/3 (pmh-dev #77).
+
+**⑤ `gate_shape_scan.sh` 여러 줄 docstring 안쪽 오탐 수리 + 판정형 질문 reps≥3 교리** (pmh-dev #76 역수확, #699): `"""`/`'''` 짝 상태를 추적해 docstring 내부를 판정 어휘 스캔에서 뺀다(selftest 8 → 10, 레인 9 → 11, L8b 되돌림 프로브 포함).
 ### [3.2.0] — 2026-09-09 — 판정 파이프라인이 실제로 돌고, 축에 이름이 붙었다
 
 **BREAKING (gate):** `scripts/**/*.py` 가 이제 HEAVY 다 — 마커 없는 파이썬 단독 커밋이 막힌다.

package/plugins/fh-meta/skills/agent-composer/SKILL.md

@@ -172,7 +172,7 @@ Default composition table by task type.
 | Plugin recommendation | plugin-recommender (S) | — |
 | Install conflict diagnosis | install-doctor (S) | — |
 | Onboarding install | install-wizard (S) — ⚠️ interactive; `--dry-run` for bg parallel | — |
-| Hub PR review | hub-cc-pr-reviewer (S) — requires PR number first | — |
+| Hub PR review | harness-pr-reviewer (S) — requires PR number first | — |
 | **Decision-maker approval review** | apex-review (S) — CTO/tech lead/QA lead personas + HTML deck | — |
 | **Project local skills** | LOCAL_SKILL_REGISTRY lookup → relevant project skill (A/S) | Per project |
 
@@ -360,7 +360,7 @@ After fan-in report, evaluate conditions and auto-suggest the next Wave:
 
 | Condition | Wave suggestion |
 |---|---|
-| ① M-tier > 0 | **Wave next-M**: fact-checker (A) → hub-cc-pr-reviewer (S) |
+| ① M-tier > 0 | **Wave next-M**: fact-checker (A) → harness-pr-reviewer (S) |
 | ② persona-innovator naming candidates > 0 | **Wave next-I**: delegate to user + asset-placement-gate (S) |
 | ③ External absorption signal High > 0 | **Wave next-E**: persona-innovator Mode E (A) + meta-prompt-builder (S) |
 | ⑤ Design conflict / 2+ conflicting suggestions | **Wave next-D**: deliberation (S) — verdict folds back into Step 4-b |

package/plugins/fh-meta/skills/auto-decorrelation/SKILL.md

@@ -229,6 +229,46 @@ format-checked `residency=(CLEAN|TAINTED|NOT_SCANNED)(...)` token on `DEGRADED_*
 `declined`. Fixtures: `scripts/test_marker_crossfamily_lanes.sh` (`r1`–`r13`). This is the SAME
 typed field Step 6 already emits into — one channel, not a second marker line.
 
+## Step 4.6 — `evidence=` : say WHAT the panel received (mandatory with `panel(...)`, 2026-09-12)
+
+`residency=` says the payload was safe to send. `evidence=` says **what was sent** — and that is the
+axis the outside world just measured as the dominant one. `arXiv:2609.10969` (48 task templates →
+2,880 scenarios, **fixed call budget** so the comparison is about the axis and not about spending
+more): a cross-model vote over **shared** evidence approved **62.9 %** of unsafe proposals against
+**22.9 %** with an **independent source** — source effect **40.9 pp** vs **11.3 pp** for model
+diversity, i.e. **3.6×**.
+
+Decide it from **how this run actually dispatched**, not from intent:
+
+| What happened | Token |
+|---|---|
+| Every member got the **same** payload (one prompt file / one diff fanned out) | `evidence=SHARED(<name the artifact>)` |
+| Each member produced its **own** evidence — own checkout, own run, own extraction | `evidence=INDEPENDENT(<what each got>)` |
+| Some of each | `evidence=MIXED(<which member got which>)` |
+
+🟥 **This skill's default flow is `SHARED`, by construction.** Step 4 builds **one** payload and fans
+it out, so a panel dispatched the ordinary way through this skill is the **62.9 % arm** — the axis it
+raises is model family, which the paper prices at 11.3 pp. That is not a reason to stop using it: a
+finder is worth its recall, and this skill's value was never precision
+(`[[feedback_decorrelation_is_a_cheap_finder]]`). It **is** a reason to stop recording that run as if
+it carried the stronger axis. Write `SHARED` and mean it.
+
+🟢 **`SHARED` is a legal, expected answer.** The hook does not block it; it blocks a `panel(...)` that
+does not say which it was, a value outside the closed three, a malformed or duplicated token, and a
+**vacuous body** on `SHARED`/`MIXED` (name the artifact — *"same staged diff"*, *"one findings.jsonl"*,
+*"identical prompt file"*). `INDEPENDENT` is self-describing and is not body-checked, deliberately:
+over-blocking the honest strong answer would train the override.
+
+**Where it lands**: the same grounds line as `residency=`, in the same `crossfamily:` field —
+`crossfamily: panel(codex) — residency=CLEAN(files=3) · evidence=SHARED(same staged diff to both) · R1, 3 findings`.
+Hook: `validate_crossfamily_leg` (grace `EVIDENCE_TOKEN_GRACE_DATE`, no retroactivity).
+Fixtures: `scripts/test_marker_crossfamily_lanes.sh` `e1`–`e10`.
+
+⚠️ **Named gap, not built today**: this skill has **no independent-source mode**. Giving it one
+(each member reconstructs the evidence itself rather than reading the author's payload) is the change
+that would move it onto the 40.9 pp axis; it is a design question, not a wiring one, and it is
+recorded as a signal rather than improvised here.
+
 🟥 **A stripped file is invisible to the reviewer, and a reviewer's default read of invisible is
 "absent," not "redacted"** (governor dogfood, 2026-09-05, same-day live use of this Step): two
 TAINTED files were stripped from a real payload and the CLEAN remainder sent cross-family — the

package/plugins/fh-meta/skills/frontier-digest/SKILL.md

@@ -215,7 +215,7 @@ Present Step 4 menu options [1]–[5]. Do not skip to [5] silently — surface t
 ## Simplification Guards
 
 - Video Tier-3 probe fails (any of `yt-dlp` / `curl_cffi` / `ffmpeg` missing, or timedtext returns 429) → fall through to operator summary; never assume `yt-dlp` works
-- If 3+ arxiv queries fail (HTTP 429): back off once → WebFetch the date-sorted `arxiv.org/list/cs.SE/recent` listing → only then HN-only. Never substitute WebSearch (recall channel, 5/5 REPEATs measured 2026-09-04); report `arxiv FAILED (429)`, never `0 items`. Detail: `SKILL_detail.md §Execution form`
+- 🟢 **arxiv transport, inverted 2026-09-12 (operator decision, condition met: 429 on 09-09·10·11·12)**: the **date-sorted `arxiv.org/list/cs.SE/recent` listing is PRIMARY**; the export API is the **fallback**; HN-only only if both fail. Same sort axis (submission date) is why the listing is promotable and WebSearch is not. 🟥 Cost of the promotion, not an oversight: the listing's ID↔title pairing is the weaker instrument, so **per-item `abs/{id}` re-resolution is now a PRIMARY-path requirement** and the progress line must name which transport produced the items. Never substitute WebSearch (recall channel, 5/5 REPEATs measured 2026-09-04); report `arxiv FAILED (429)`, never `0 items`. 🟥 **On the listing path, pair ID↔title INSIDE one list entry, never by zipping two ordered lists — measured 2026-09-12: 51 IDs vs 50 titles, so pairs drift partway down and mispair while looking well-formed (3 items caught, 0 shipped). Until that is what happened, per-item `abs/{id}` re-resolution is MANDATORY and a title mismatch DROPS the item.** Detail: `SKILL_detail.md §Execution form`
 - On curl timeout, skip that item and continue with the rest
 - If synthesis result exceeds 400 characters, retain top 3 items and truncate the rest
 - Without `--save`, do not create files (conversation output only)

package/plugins/fh-meta/skills/frontier-digest/SKILL_detail.md

@@ -28,14 +28,52 @@ load: on-demand
 > fell back to **WebSearch**, and the 09-04 run measured what that does: 5/5 papers it returned were
 > REPEATs. Search sorts by *canonicity*, the export API by *submission date* — a search fallback is
 > a **recall channel, not a discovery channel**, and swapping it in silently converts «unrun» into
-> «nothing new» (`not found ≠ 0`). Order on 429:
-> 1. back off once (≥30s), retry the same query;
-> 2. still 429 → WebFetch the **date-sorted HTML listing** `https://arxiv.org/list/cs.SE/recent`
+> «nothing new» (`not found ≠ 0`).
+>
+> 🟢 **TRANSPORT ORDER INVERTED 2026-09-12 (operator decision) — the listing is now PRIMARY.**
+> Condition stated by the operator was *«if the errors keep coming»*, and they did, measurably:
+> `429` appears in the digest record on **09-09 · 09-10 · 09-11 · 09-12 — four consecutive days**,
+> past the `N≥3` bar twice over. Paying two failed calls plus a backoff every single day to reach
+> items that the listing reaches anyway is not a fallback, it is a toll. New order:
+> 1. **PRIMARY — WebFetch the date-sorted HTML listing** `https://arxiv.org/list/cs.SE/recent`
 >    (and `https://arxiv.org/list/cs.AI/recent` if the first is empty of agent/harness items), keep
->    items whose title/abstract match the three query phrases, cap 6 — same sort axis as the API;
-> 3. only if that also fails → HN-only, and the progress line says `arxiv FAILED (429)`, never `0 items`.
+>    items whose title/abstract match the three query phrases, cap 6. **Same sort axis as the API**
+>    (submission date) — that property is *why* this one is promotable and WebSearch is not.
+> 2. **FALLBACK — the export API query** (back off ≥30s between attempts). Kept on the ladder on
+>    purpose: if arXiv redesigns the listing markup, the channel must not disappear with it.
+> 3. only if both fail → HN-only, and the progress line says `arxiv FAILED (<which transports, why>)`,
+>    never `0 items`.
+>
+> 🟥 **What this costs, stated rather than discovered later.** The listing path has the **weaker
+> instrument** — its ID↔title pairing is the defect documented immediately below, and promoting it
+> moves that defect from the fallback onto the **main** path. Therefore the two rules below are no
+> longer conditional-on-being-in-fallback: **per-item `abs/{id}` re-resolution is a PRIMARY-path
+> requirement**, one fetch per shipped ID, and a title mismatch drops the item. That is a real
+> per-run cost increase (≈N extra fetches for N shipped items) and it is the price of the promotion,
+> not an oversight. The old order paid 2 failed calls + a backoff instead; this pays N verified reads.
+> 🟥 The progress line must now name **which transport produced the items** (`arxiv via listing` /
+> `arxiv via export-api`) — otherwise a reader cannot tell which instrument's limits apply to the run.
 > 🟥 WebSearch is **not** on this ladder. If a run used it anyway, the arxiv leg is reported as
 > `RECALL-ONLY (WebSearch)` and its items are excluded from the NEW/REPEAT count.
+>
+> 🟥 **On the listing path, do NOT pair ID and title by position — pair them inside one block
+> (measured 2026-09-12, this runner, and it is the phantom class this skill already has on record).**
+> The listing yielded **51 IDs and 50 titles**, so two parallel extractions drift partway down and
+> every pair after the drift point is wrong *while looking well-formed*. Concretely measured that run:
+> `2609.10550` was attached to *"When Passing Tests Hides Vulnerabilities"* by the positional parse,
+> but its own `abs` page is *"Optimizing AI Inference Across the Deployment Stack"* (2026-07-01);
+> `2609.11060` and `2609.11008` mispaired the same way. 3 items dropped, 7 verified, **0 shipped
+> unverified** — it was caught only because every shipped ID was re-resolved.
+> **Two rules, and the second is the floor:**
+> ① read each entry as **one unit** (the ID and the title that sit in the *same* list item), never as
+>   two ordered lists zipped together. A count mismatch between the lists is not a warning to
+>   reconcile — it is proof the zip is already wrong.
+> ② until ① is demonstrably what happened, **per-item `abs/{id}` re-resolution is MANDATORY on this
+>   path**, not optional — one fetch per shipped ID, confirming the title. Any ID whose `abs` title
+>   disagrees is **dropped**, never shipped with a note.
+> ⚠️ There is no hook here and none is possible: the parse happens inside the session that read the
+> page, so this is salience over a manual floor. The floor that actually held on 2026-09-12 was ② —
+> keep it even after ① looks solved (`[[feedback_anchor_can_be_decorative]]`).
 
 ### HackerNews (Algolia API)
 

package/plugins/fh-meta/skills/{hub-cc-pr-reviewer → harness-pr-reviewer}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: hub-cc-pr-reviewer
+name: harness-pr-reviewer
 description: Checks a submitted PR against the environment's baseline assets (CLAUDE.md, memory, naming, asset classification) and attaches a review comment with a merge recommendation. 5 steps — diff read, 8-area consistency check, self-catch, comment, merge recommendation.
 user-invocable: true
 allowed-tools: ["Bash", "Read", "Grep", "Glob"]
@@ -15,13 +15,21 @@ complexity_routing:
 
 > **Note:** The original developer is the forge-harness original developer (development source + meta-monitoring home). In external user install environments, the install environment user themselves is the baseline integrity gate operator (following path B generalization baseline / `SKILL_detail.md §External User Environment Adaptation Path` §).
 
-# hub-cc-pr-reviewer — Hub Gate Operation Rule Automation
+# harness-pr-reviewer — Hub Gate Operation Rule Automation
 
 When a PR is submitted, checks consistency against the user environment's baseline assets (CLAUDE.md · memory · naming · asset classification) and attaches a review comment. 5-step: diff read → 8-matrix check → self-catch → comment attachment → merge recommendation.
 
 ## Activation Triggers
 
-1. **PR #N input**: *"Review PR #N"* / *"Check PR #N"* / *"hub review"* / *"baseline consistency check"*
+> **Renamed from `hub-cc-pr-reviewer` (2026-09-12, operator decision).** The old name said where it
+> was born (the hub's own cc reviewing its own PRs). What it actually is — measured on qasp/pmh PRs in
+> 2026-09 — is the **standalone doorway through which a field harness verifies its own submitted PR
+> with FH's review capability**. Old-name utterances ("hub-cc-pr-reviewer", "hub review", "hub cc
+> review") still route here; downstream forks that carry the old directory name (PMH) keep working
+> until they sync. No redirect stub directory is shipped (same policy as `phantom-quench`).
+
+
+1. **PR #N input**: *"Review PR #N"* / *"Check PR #N"* / *"hub review"* (old-name alias) / *"harness PR review"* / *"baseline consistency check"*
 2. **Action leader cc → hub sync point**: Large decision area PR catch (following Option C Hybrid policy — memory creation / CLAUDE.md change / CATALOG round / skill v0.x evolution / policy change / asset synergy branch judgment)
 3. **Hub cc session entry**: Layer A auto-read recent external commit catch (auto-discover new PRs)
 
@@ -75,6 +83,54 @@ Self-precision catch areas after first cc review (following previous PR self-cat
 
 Self-catch areas 0 items = skip this entire catch matrix — do not pad with token-filling to make the section look populated.
 
+🟥 **Self-catch is a CUE, not a verdict.** Its output is the *input* to Step 3.5 below. A self-catch
+that finds nothing does not clear Axis 2 or Axis 3 — measured 2026-09-08/11: a reviewer session with
+no external cue ran the load-bearing gate **0/51**, and a PMH session reviewing qasp PR #13 dispatched
+the cross-family panel only after the operator pushed (pmh-dev #77). The self-check *is* the reviewer;
+it cannot be its own decorrelation.
+
+### Step 3.5. Axis 2 · Axis 3 — mandatory dispatch lane (mechanical trigger, typed degrade)
+
+Axis 1 above is wired at mandatory strength (`--pr` mode, typed verdict). Until 2026-09-12 Axis 2
+(adversarial panel) and Axis 3 (phantom-quench) had **no dispatch block in this skill at all** — only
+the self-catch matrix — so they fired only when someone said so. This step closes that with a trigger
+that is decided from the PR, never from the reviewer's feel.
+
+**Trigger — mandatory if ANY of these holds** (compute all three; record which fired):
+
+```
+(i)   Step 2 matrix has ≥1 ❌  (Inconsistent)
+(ii)  the PR touches a load-bearing FH asset: SKILL.md · SKILL_detail.md · .claude/rules/*.md ·
+      knowledge/shared/rules/*.md · templates/** · scripts/**/*.sh · scripts/**/*.py ·
+      plugins/*/agents/**   (same list as CLAUDE.md §FH Improvement 4-Axis Auto-Gate)
+      → gh pr diff "$PR" --name-only | grep -E '(^|/)SKILL(_detail)?\.md$|^\.claude/rules/|^knowledge/shared/rules/|^templates/|^scripts/.*\.(sh|py)$|^plugins/[^/]+/agents/'
+(iii) a merge recommendation (Step 5) will be issued for this PR
+```
+
+None of the three → record `axis2: not-triggered(reason)` · `axis3: not-triggered(reason)` in the
+comment capsule and continue. This is the *only* non-run path, and it is stated, never silent.
+
+**Axis 2 — cross-family adversarial panel.** Dispatch **`auto-decorrelation`** on the PR diff (reuse,
+not a new engine — its Step 1 load-bearing predicate, Step 4.5 residency screen and Step 6 degrade
+ladder all apply unchanged). What this step owns is only the *decision to dispatch*. The result lands
+as the typed `crossfamily:` value: `panel(<families>)` with the source-grounded findings, or one of
+`declined` · `DEGRADED_SINGLE_FAMILY` · `DEGRADED_PANEL_UNUSED` · `UNKNOWN` **with grounds** — e.g.
+consent OFF, no different-family sidecar reachable, quota exhausted. A degrade value is allowed; an
+*absent* value is not. 🟥 A judgment-type question put to the panel (by-design vs fail-open, is this
+a regression) needs **reps ≥ 3**; a 1-of-3 split is recorded as *unresolved*, never as CONCUR
+(`field_verdict_crossfamily_gate.md` §4-2).
+
+**Axis 3 — phantom-quench.** Run `/phantom-quench` over the PR's changed documentation surfaces
+(any `*.md` in the diff, plus every path/citation the PR body asserts). Mechanical N/A only when
+`gh pr diff --name-only` has zero `*.md` files **and** the PR body cites no path — record
+`axis3: N/A(no doc surface)`. A phantom finding is a ❌ in the comment capsule, and it flips the Step 5
+recommendation the same way an Axis 1 M-tier block does.
+
+**Degrade direction.** This is a review surface (reversible): tooling-down does not block the
+*comment*, but the merge recommendation must carry the degrade value verbatim and read
+**NOT-CONVERGED** while Axis 2 is a `DEGRADED_*`/`UNKNOWN` value on a load-bearing PR. Silent
+same-family pass is the one outcome this step forbids.
+
 ### Step 4. Review Comment Attachment
 
 **Mandatory before any `gh pr comment`: run `/public-surface-audit` over the composed comment text.**
@@ -176,6 +232,22 @@ All 5 Steps completed
     passes; they mean Axis 1 did not examine this PR and the recommendation
     may not cite it as green
 
++ Step 3.5 trigger computed from the PR (matrix ❌ count · load-bearing path
+  grep · merge-recommendation requested) and recorded with which clause fired
+  — measured: the three booleans appear in the comment capsule. Absent = the
+    step did not run, which is the pre-2026-09-12 defect (pmh-dev #77)
+
++ Axis 2 dispatched via auto-decorrelation when triggered, landing a typed
+  `crossfamily:` value (panel(...) or a DEGRADED_*/UNKNOWN value WITH grounds)
+  — mandatory-pass: the value exists and is one of the closed enum. A
+    load-bearing PR whose value is DEGRADED_*/UNKNOWN leaves the merge
+    recommendation NOT-CONVERGED; judgment-type panel questions carry reps ≥ 3
+    or are recorded unresolved
+
++ Axis 3 phantom-quench run over the PR's doc surfaces when triggered, or
+  N/A(no doc surface) decided by `--name-only` grep, never by feel
+  — mandatory-pass: a phantom finding is a ❌ that flips the recommendation
+
 + /public-surface-audit run over the composed comment text BEFORE any
   gh pr comment
   — mandatory-pass, fail-closed: CLEAN attaches; REVIEW/LEAK redact-and-rescan;

package/plugins/fh-meta/skills/{hub-cc-pr-reviewer → harness-pr-reviewer}/SKILL_detail.md

@@ -1,6 +1,6 @@
 ---
-name: hub-cc-pr-reviewer-detail
-description: On-demand detail for hub-cc-pr-reviewer — step bash commands, comment template, sister-asset utilization, external-environment adaptation, disable path, and persona synergy handling. Read when executing a step or operating in an external/own-PRS/deep-insight environment.
+name: harness-pr-reviewer-detail
+description: On-demand detail for harness-pr-reviewer — step bash commands, comment template, sister-asset utilization, external-environment adaptation, disable path, and persona synergy handling. Read when executing a step or operating in an external/own-PRS/deep-insight environment.
 load: on-demand
 ---
 

package/plugins/fh-meta/skills/harvest-loop/SKILL_detail.md

@@ -35,7 +35,7 @@ Meaning of isolation: The Critic reads the synthesizer conclusion but does not i
 - FAIL after re-synthesis → auto-persist as `fh_signal` on hold (no additional retries)
 - Maximum retries: **1**
 
-**Post-Core-Skill Critic Verdict Connection**: Following core skills can have Critic called inline after completion: harness-doctor · verify-bidirectional · hub-cc-pr-reviewer · context-doctor · sim-conductor. Trigger: immediately after completion announcement + "steel-quench" / "re-validate" / "run Critic" utterance.
+**Post-Core-Skill Critic Verdict Connection**: Following core skills can have Critic called inline after completion: harness-doctor · verify-bidirectional · harness-pr-reviewer · context-doctor · sim-conductor. Trigger: immediately after completion announcement + "steel-quench" / "re-validate" / "run Critic" utterance.
 
 ---
 
@@ -150,7 +150,7 @@ grep -rl "complexity_routing" plugins/*/skills/*/SKILL.md
 
 # 2. Aggregate escalation records from fh_signal files
 grep -rh "" tracks/_meta/fh_signal_*.md 2>/dev/null | \
-  grep -oE "(harness-doctor|verify-bidirectional|hub-cc-pr-reviewer|context-doctor|sim-conductor|agent-composer|harvest-loop|steel-quench)" | \
+  grep -oE "(harness-doctor|verify-bidirectional|harness-pr-reviewer|context-doctor|sim-conductor|agent-composer|harvest-loop|steel-quench)" | \
   sort | uniq -c | sort -rn
 ```
 

package/plugins/fh-meta/skills/install-doctor/SKILL.md

@@ -169,7 +169,7 @@ grep -i -n "pr\|pull request\|audit\|review\|weekly" CLAUDE.md 2>/dev/null | hea
 ```
 
 Judgment:
-- Existing PR convention present → possible priority conflict with `hub-cc-pr-reviewer` ⚠️
+- Existing PR convention present → possible priority conflict with `harness-pr-reviewer` ⚠️
 - Existing weekly audit present → possible format conflict with `harvest-loop` ⚠️
 
 ### 2-2. Skill Trigger Conflicts

package/plugins/fh-meta/skills/install-wizard/SKILL.md

@@ -322,7 +322,7 @@ Only activate the cluster matching the utterance — loading all skills degrades
 | **A — Harvest/Evolution** | harvest · session wrap-up · pattern · reverse absorption · fh evolution | field-harvest · harvest-loop · contention-layer · verify-bidirectional |
 | **B — Diagnosis (doctor)** | diagnose · health · check · inspect · token waste · install conflict · doctor | harness-doctor · context-doctor · sim-conductor · install-doctor · install-wizard |
 | **C — Compose (composer)** | agent composition · parallel · compose prompt · context card | agent-composer · meta-prompt-builder |
-| **D — Audit/Review** | review · audit · PR review · steel quench · adversarial · lint · placement | hub-cc-pr-reviewer · steel-quench · apex-review · harness-doctor (--lint) · marketplace-gate · asset-placement-gate |
+| **D — Audit/Review** | review · audit · PR review · steel quench · adversarial · lint · placement | harness-pr-reviewer · steel-quench · apex-review · harness-doctor (--lint) · marketplace-gate · asset-placement-gate |
 | **E — Explore/Frontier** | trends · plugin recommendation · synergy · frontier | frontier-digest · plugin-recommender · cross-ecosystem-synergy-detection |
 | **F — Common (always-on)** | — | convergence-loop · deliberation (fh-commons) |
 

package/plugins/fh-meta/skills/meta-prompt-builder/SKILL.md

@@ -53,7 +53,7 @@ When Goal is provided in natural language (e.g., "I need to report to the team l
 | "analyze", "diagnose", "something seems off" | harness-doctor → context-doctor | Structural + contextual diagnosis simultaneously |
 | "simulate", "validate", "meta" | sim-conductor (D-code or Area B) | Multi-perspective validation |
 | "install", "setup", "onboarding" | plugin-recommender → install-wizard | Recommend then install |
-| "review", "code check" | sim-conductor D-code → hub-cc-pr-reviewer | Code review chain |
+| "review", "code check" | sim-conductor D-code → harness-pr-reviewer | Code review chain |
 
 Output format:
 ```

package/plugins/fh-meta/skills/pipeline-conductor/SKILL.md

@@ -246,7 +246,7 @@ Resume is scope-bound — the scope from Step 0 is preserved across resume calls
 After a `CLEAN (--full)` or `PENDING` sweep, the following are natural follow-ons:
 
 - `field-harvest` — harvest patterns surfaced during the sweep
-- `hub-cc-pr-reviewer` — if sweep was run pre-PR, feed results into PR review
+- `harness-pr-reviewer` — if sweep was run pre-PR, feed results into PR review
 - `agent-composer` — if multiple fix tasks are needed across the Pending list, compose agents to resolve them in parallel
 - `return-path-gate --all` — if Step 0.5 surfaced OPEN chains, run a full chain closure audit
 

package/plugins/fh-meta/skills/plugin-recommender/SKILL.md

@@ -98,7 +98,7 @@ When queried for a specific capability (e.g., "adversarial reviewer for bash cod
 ⚠️ **A failed or empty discovery lane is NOT "no candidates".** Both CLIs above list only *configured* marketplaces, and a non-zero exit / empty array means the lane did not answer — not that nothing exists. Report each lane's state explicitly (`EXECUTED` / `EMPTY` / `FAILED: <stderr>`) and never render a `FAILED` lane as a zero result; a lane that could not run must be re-run or replaced by the web-search fallback (Priority 3) before you tell the user nothing was found.
 
 **Discovery priority**: built-in (Tier 0) > installed > FH native > Tier 1 (any platform) > Tier 2 > Tier 3 > Tier 4
-**Tier 0 guard**: FH native wins over a built-in only when the FH skill adds governance the built-in lacks (e.g. `/goal` → `goal-quench` adds budget+quality gates; code diff review stays with built-in `/code-review`, FH-asset coherence with `hub-cc-pr-reviewer`)
+**Tier 0 guard**: FH native wins over a built-in only when the FH skill adds governance the built-in lacks (e.g. `/goal` → `goal-quench` adds budget+quality gates; code diff review stays with built-in `/code-review`, FH-asset coherence with `harness-pr-reviewer`)
 
 **When sim-conductor chains here for persona discovery**: apply the same platform-aware search scoped to persona/simulation/review capability tags. Return discovered agents with their Tier rating so sim-conductor can decide whether to install or use a built-in brief.
 

package/plugins/fh-meta/skills/sim-conductor/SKILL.md

@@ -542,7 +542,7 @@ Verdicts: PASS · CONDITIONAL_PASS (S/R only, or Area B cadence skip) · FAIL (M
 
 ## Operations Notes
 
-**AI-AI Loop Bias Defense**: sim-conductor → hub-cc-pr-reviewer → sim-conductor loop shares LLM cognitive blind spots. Internal convergence is "provisional convergence" — elevated only by external review or human gate. Area A/B convergence without human gate = incomplete.
+**AI-AI Loop Bias Defense**: sim-conductor → harness-pr-reviewer → sim-conductor loop shares LLM cognitive blind spots. Internal convergence is "provisional convergence" — elevated only by external review or human gate. Area A/B convergence without human gate = incomplete.
 
 ---
 

package/plugins/fh-preprep/.claude-plugin/plugin.json

@@ -0,0 +1,22 @@
+{
+  "name": "fh-preprep",
+  "version": "3.5.0",
+  "engines": {
+    "claudeCode": ">=1.0.0"
+  },
+  "description": "preprep — a presentation-preparation harness. One CLI over 17 lanes (7 built-in + 10 module) that reads a declared surface manifest and checks a talk's assets against each other: canonical-term ledger, jargon glosses, retired vocabulary, inter-slide dependencies, pacing, get/put round-trips, adjacent duplication, slide relations, geometry, attribute consistency, fonts, diagrams, slide refs, and screen-parity between manuscript and deck. Verdict contract: exit 0 pass / 1 findings / 2 cannot-judge — a surface that could not be read is UNMEASURED, never 0. Every lane ships a known-pair fixture.",
+  "author": {
+    "name": "chrono-meta",
+    "email": "chrono-meta@users.noreply.github.com"
+  },
+  "keywords": [
+    "presentation",
+    "deck-review",
+    "ooxml",
+    "pptx",
+    "terminology",
+    "screen-parity",
+    "known-pair",
+    "verdict-contract"
+  ]
+}