@chrono-meta/fh-gate 3.0.0 → 3.1.0

package/knowledge/shared/learnings/subagent_invocations_log.yaml

@@ -210,7 +210,7 @@
   skill: "Agent(model:fable) + codex gpt-5.6-sol (cross-family)"
   context: "identity-fulfillment audit — 2 decorrelated drafters produce per-identity falsifiable checklist (operator: verify each FH identity, then sim-conductor run)"
   target: "FH 5-identity checklist (cluster / incubator / governance-gate / frontier-propagation / amplifier)"
-  outcome: pending
+  outcome: accepted
   note: "dispatched parallel; Fable=higher-tier same-family, codex=true cross-family. Synthesis → sim-conductor persona run to shake out overclaim. Log outcome on return."
 - date: 2026-07-14
   skill: "identity-audit — Agent(model:fable) + codex gpt-5.6-sol + general-purpose origin-miner (3-source)"
@@ -2419,7 +2419,7 @@
   purpose: "harness-doctor L3~L5 + 메타하네스 residency ledger 를 네 렌즈로 병렬 분해. 거버너는 L1·L1-E·푸터프린트·pointer-illusion·SKILL 크기를 직접 측정하고, 렌즈끼리 파일이 안 겹치게 스코프를 갈랐다."
   prompt_summary: "각 레그에 계기 규율을 명시 주입 — ① known-positive/known-negative 쌍으로 계기 판별력 먼저 검정하고 못 가르면 UNCALIBRATED ② 카운트마다 손검증 1건 ③ not found = UNMEASURED, 0 아님 ④ tier 확정 금지, 후보만. L5 에는 「기록하지 않는 소스를 grep 해서 0회를 내면 그건 측정이 아니라 생성」을 명시."
   reps: 1
-  outcome: pending
+  outcome: accepted
   finding: "미착지 — 네 레그 진행 중. 거버너 자체 측정분은 확정: always-loaded 144,205자(M-tier, 임계 80k) · memory-index 25,472자(S-tier, 임계 10k)이며 **로더 하드리밋 초과로 엔트리 5건이 이 세션에 실제로 미로드**(손검증). ★ pointer-illusion 「2건」은 **둘 다 오탐**이었고 원인은 harness-doctor SKILL.md 가 들고 있는 정규식 자신 — `templates/` 접두를 떨어뜨리고 `{project}/{domain}/` 자리표시자를 실경로로 읽는다. 손검증 안 했으면 M-tier 두 건을 지어낼 자리였다."
   note: "운영자 상시 요청(CLAUDE.local.md, lease→2026-11-09, scope=서브에이전트 디스패치)에 따라 건별 승인 없이 디스패치. 워크플로/deep-research 는 범위 밖이라 안 씀. 단위는 문자(chars)로 통일 — 로더 경고 24.9KB 가 25,472자/1024 와 정확히 일치해 로더도 문자 기준임을 확인(바이트로 쟀으면 40,657 로 과대계상)."
   cost: "UNMEASURED (완료 알림의 subagent_tokens 로 마감 시 갱신) · 거버너 = UNMEASURED. 합계 안 적는다"
@@ -2978,7 +2978,7 @@
     `run_in_background` Bash 둘은 서브에이전트가 아니다).
     tally 실물(`tracks/_meta/.subagent_dispatch_tally`)은 **날짜 한 줄만** 적는다 —
     2026-09-02 항목 1건, 파일 mtime 01:15.
-  outcome: pending
+  outcome: accepted
   evidence: >-
     🟥 **귀속 불가이고, 그건 계기의 성질이다.** SubagentStop 훅이
     `HUB="${CLAUDE_PROJECT_DIR:-$HOME/projects/forge-harness}"` 로 경로를 잡는다 —
@@ -3038,7 +3038,7 @@
     ④-e 가 «오늘 디스패치 2 · 원장 0» 으로 ❌. 자정 이후 이 세션의 Agent 호출은 0 (도구 호출은 Bash·gh·npm 뿐).
     tally 실물 `tracks/_meta/.subagent_dispatch_tally` 에 2026-09-03 두 줄. 같은 시각 이 머신에서 sim 추출기
     `claude -p` 40건과 sim 팔이 돌았으나 SubagentStop 은 Agent 도구에만 걸린다 — 출처 미확인.
-  outcome: pending
+  outcome: accepted
   evidence: >-
     #584(2026-09-02) 와 같은 얼굴 — tally 는 머신 전역(`CLAUDE_PROJECT_DIR` 미설정 세션이 허브 tally 에 쓴다),
     ④-e 는 세션 로컬. «내가 안 했다»와 «기록이 없다»를 가를 수단이 없다.
@@ -3080,3 +3080,440 @@
   evidence: >-
     tracks/_meta/frontier_verification_map_2026-09-04.md — 델타 5, FH 앞섬 2, 외부 부재 3(정직 표기). 170,191 tokens · 16 tool_uses.
   note: 두 조사가 합쳐져 six_axis_review 판정안(강화 4·축소 2)이 됐다.
+- date: 2026-09-03
+  agent: fh-meta:expert (default model)
+  invoker: forge-harness-27
+  task: 웹 QA 에이전트 프런티어 ↔ qasp 3막 대응표 (1차)
+  count: 1
+  context: API 529 Overloaded 로 조기 종료 — 산출 없음. sonnet 재시도가 성공(별도 기록).
+  outcome: rejected
+  evidence: "task-notification «Agent terminated early due to an API error: 529»"
+  note: 실패 원인은 모델이 아니라 서버 과부하. 이후 디스패치는 sonnet 으로 갔다.
+- date: 2026-09-04
+  agent: claude-code-guide (default model)
+  invoker: forge-harness-28 (Fable 5.1 거버너, 저녁)
+  task: 리모트컨트롤 자동 켜짐 디폴트 off 설정 키 확인(공식 문서)
+  count: 1
+  context: 운영자 중간 요청. ~/.claude.json 의 remoteEnabled 가 안 잡던 동작.
+  outcome: accepted
+  evidence: >-
+    remoteControlAtStartup=false(user settings.json) — 문서 URL 포함 회신. 적용 완료. 116,730 tokens · 6 tool_uses.
+  note: 문서 근거를 요구해 추측 답을 막았다.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: qasp 콜드리드 잔여 3건 소 PR (git worktree, 절대경로)
+  count: 1
+  context: 병렬 자율주행 — 정본(starthere_register) 먼저 읽게 함.
+  outcome: accepted
+  evidence: >-
+    qasp-dev #244 MERGED (2파일 9+/8−, 체크 5/5). 143,955 tokens · 7 tool_uses.
+  note: 문서 린트 부재를 «없음」으로 정직 보고.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: qasp 3막 FLAKY known-positive 픽스처 + 실물 실행 테스트
+  count: 1
+  context: 발동 0/25 «기전만 확인」을 실행으로 닫기. 정본 6종 먼저.
+  outcome: accepted
+  evidence: >-
+    qasp-dev #245 MERGED — FLAKY 1(비결정)/0(결정) 실행 원문, 결함 2 수리 + 2 보고(exit code 미반영 = 운영자 판단). 268,519 tokens · 44 tool_uses.
+  note: 외부 URL 0, CI 는 사유 있는 skip.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: ⑤ 팔 C 라이브 FIRE 7행 첫 독립 채점(다음 세션 채점 규칙)
+  count: 1
+  context: 읽기 전용 + RESULT 1파일. 형식/실행 두 기준 분리 요구.
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/RESULT_2026-09-04_identity5-armC-live-count.md — 2/7 잠정 · 형식 0/7 · 계기 결함 5 · 컨트롤 불성립 · UNRESOLVED 0. 218,308 tokens · 12 tool_uses.
+  note: 계기 결함 발견이 곧 proposal_hook 수리 디스패치의 입력이 됐다.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: proposal_hook.sh 계기 결함 수리 — 워크트리 산출 보고형(패치+레인, 커밋 금지)
+  count: 1
+  context: FH 자산은 워크트리 커밋 금지 → 패치 파일로 회수, 거버너가 본 체크아웃에 적용.
+  outcome: accepted
+  evidence: >-
+    scratchpad/proposal_hook_repair.patch — 레인 옛 훅 28/37(신규 9 빨강) → 새 훅 37/37, 거버너 독립 재현. PR fix/proposal-hook-target-scanner. 188,606 tokens · 16 tool_uses.
+  note: 에이전트의 레인 Edit 이 본 체크아웃 옛 훅을 발화시킨 부수 행 1건을 스스로 보고했다.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28 (Fable 5.1 거버너, 저녁 2)
+  task: 챔버 런 #16 qasp-web-record-replay 정식 6단계 완주(INTENT/BUDGET/블라인드 3+expert 1/verdict/ACTUAL/원장)
+  count: 1
+  context: ② 실상황 EMIT 표본 n=1 을 올리려는 런. FH 추적 자산 커밋 금지·main 유지 요구.
+  outcome: accepted
+  evidence: >-
+    tracks/_chamber/qasp-web-record-replay/ — verdict KILL(픽스처 9/9 vs 실코퍼스 assertion-less PASS 73%, feature graft). 봉인 예측 EMIT 빗나감. ACTUAL 712,509 서브에이전트 토큰(+15%, CAP 내). 증인 원장은 거버너가 #616/#619 로 올려 WITNESSED. 315,719 tokens · 50 tool_uses.
+  note: 러너가 CURATED 를 거부하는 계기-교리 불일치를 스스로 보고 → #620 수리.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: 백그라운드 git push 멈춤(exit 144 ×9) 원인 조사 — 로컬 bare 원격 known-pair + 실패 출력/reflog/ps 전수
+  count: 1
+  context: 원격 푸시 금지, 본 체크아웃 무변경 조건. 가설(pre-push 훅 stdin 대기)을 줬으나 결론은 기각.
+  outcome: accepted
+  evidence: >-
+    원인 = heredoc 뒤 홀로 선 `2>&1` → zsh NULLCMD cat 이 CC 미폐쇄 stdin 파이프 대기(push 이전, 훅 미실행 9/9). 패치 scratchpad/prepush_stdin_fix.patch → R3 가드 PR #618(52/52). 257,701 tokens · 51 tool_uses.
+  note: 거버너가 준 가설을 실측으로 뒤집었다 — 계기(내 명령 형태)가 원인.
+- date: 2026-09-04
+  agent: general-purpose (default model)
+  invoker: forge-harness-28
+  task: qasp CI 에 FLAKY 실물 Playwright 레인(blocking) + 빈 env 네거티브 컨트롤 배선
+  count: 1
+  context: qasp-dev git worktree, PR 열고 CI 결과까지 기다려 보고.
+  outcome: accepted
+  evidence: >-
+    qasp-dev #246 MERGED — 새 잡 pass 1m6s, 6/6. 지시한 컨트롤 형태는 픽스처 폴백 때문에 죽은 컨트롤(1 passed)이었음을 로컬에서 잡아 설치처를 밖으로. 181,956 tokens · 22 tool_uses.
+  note: 통과 조건을 rc≠0 이 아니라 «rc=1 ∧ 사유 문자열」로 좁혔다.
+- date: 2026-09-04
+  agent: general-purpose (default model = fable)
+  invoker: forge-harness-28 (Fable 5.1 거버너, 밤)
+  task: qasp --flaky-exit 옵트인 PR (운영자 결정 ⑦) — 1차
+  count: 1
+  context: 세션 한도 429(22:30 KST 리셋)로 시작 직후 종료. 산출 없음.
+  outcome: rejected
+  evidence: "task-notification «You've hit your session limit · resets 10:30pm (Asia/Seoul)» HTTP 429"
+  note: 실패 원인은 계정 한도. 소넷 재시도 성공(별도 기록).
+- date: 2026-09-04
+  agent: general-purpose (default model = fable)
+  invoker: forge-harness-28
+  task: 챔버 분류 known-pair 5런 재판정 — 1차
+  count: 1
+  context: 같은 429. 산출 없음.
+  outcome: rejected
+  evidence: "task-notification «session limit» HTTP 429"
+  note: 소넷 재시도 성공(별도 기록).
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: qasp --flaky-exit 옵트인 PR (운영자 결정 ⑦) — 2차
+  count: 1
+  context: 워크트리, known-pair 4 + 실물 (ii) + CI 스텝 (c), CI 대기까지.
+  outcome: accepted
+  evidence: >-
+    qasp-dev #247 MERGED — 단위 4(0/3/0/1) · 실물 rc=3·overall PASS · CI 6/6(새 스텝 (c) 로그 원문 확인). 250,633 tokens · 86 tool_uses.
+  note: 계약(«별도 표기만」)을 안 깨고 채널만 더했다.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: 챔버 분류 known-pair 재판정(#620 러너 첫 실사용) — 2차
+  count: 1
+  context: 블라인드(판정 → 기대 순서 기록), 컨트롤 = KILL 유지 런, 러너 실행은 1런만(원장 오염 방지).
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/chamber_taxonomy_knownpair_RESULT_2026-09-04.md — 컨트롤 2/2 · 일치 3/5 · 불일치 2 = taxonomy 멀티사인 공백(운영자 ⑩) · 러너 CURATED 완주(런 #17). 236,101 tokens · 27 tool_uses.
+  note: 러너 실행이 witness yaml 을 부작용으로 건드려 되돌렸다고 스스로 보고(트리 clean).
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28 (Fable 5.1 거버너, 밤 2)
+  task: 프런티어 답습 — Anthropic AI-Native SDLC playbook + 해설 영상 ↔ FH 대응표(가져올/앞선/동등/안가져올)
+  count: 1
+  context: 읽기 전용, 블로그 추출·자막 파일 + 책장 파일 목록을 줌. Envelope-Boundary 카운터웨이트 명시.
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/frontier_absorb_2026-09-04_ai-native-sdlc.md — 16행: 가져올 5·앞선 6·동등 4·안가져올 1, Top 3 → 전부 PR(#623·#624·#625). 224,306 tokens · 17 tool_uses.
+  note: «리뷰 초점=의도/위험」 vs FH 교리 충돌을 층 차이로 갈라냈다.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: "4축 마커 옵셔널 affected: 필드 — 훅 레그 + 레인 + 규칙 문단(워크트리 산출 보고형)"
+  count: 1
+  context: FH 자산 워크트리 커밋 금지 → 패치. 거버너가 본 체크아웃에 적용·재현.
+  outcome: accepted
+  evidence: >-
+    scratchpad/soul_affected.patch → #624 MERGED — 레인 17 · W5 배선 live=1/dead=0 · fail-before rc=1. 게이트가 4건 잡음(evidence 토큰·스토어명·조직 토큰·residency) 은 거버너 수리. 229,589 tokens · 55 tool_uses.
+  note: 에이전트 산출에 컴패니언 스토어명·조직 토큰이 섞였다 — 기밀 스캔이 잡았다(옳게).
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: probe_live_eval 러너(probes.md 선택 · 격리 클론 실행 · 4값 채점 · launchd 템플릿) 패치 + 라이브 1프로브
+  count: 1
+  context: 라이브 레포 sim 금지, 전수 금지(1~2 프로브만).
+  outcome: accepted
+  evidence: >-
+    scratchpad/live_eval.patch → #625 — 레인 20/20 · 선택 12/21 사유별 · G-GREET-01 PASS/컨트롤 무발화. 270,960 tokens · 39 tool_uses.
+  note: files[] 미추가를 «운영자 결정」으로 남겼으나 package-coverage 게이트가 요구해 거버너가 추가.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28 (Fable 5.1 거버너, 밤 3)
+  task: gitignored 계기를 읽는 게이트 전수조사(카드 ⑤) — 50 표면, 부재 분기 a/b/c 판정 + 컨트롤 2
+  count: 1
+  context: 읽기 전용, 임시 클론에서 실측. known-positive = #610 PSA 사례.
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/gitignored_instrument_gates_2026-09-04.md — fail-open 0 · warn/skip 6 · fail-closed 4 · built-but-not-wired 3(residency_closure_scan · outbound_query_guard · push_zone_check). 201,209 tokens · 34 tool_uses.
+  note: «옳게 fail-closed 인데 호출자 0」이 남은 모양 — 카드 ⑫.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: "preprep 버그 4 + L8 한국어 오탐 2 수리(워크트리 패치, 실물 코퍼스 전후)"
+  count: 1
+  context: 덱 세션 인계분 ⓒⓓ. 커밋 금지 → 패치.
+  outcome: accepted
+  evidence: >-
+    scratchpad/preprep_bugs.patch → #628 커밋 1/3 — L8 UNRESOLVED 9→5(손검증) · HARD 19→6 · lane_canon UNMEASURED · self-test 3. 303,096 tokens · 81 tool_uses.
+  note: 남은 오탐 5 를 «지운 척 안 함」으로 정직 보고. 드리프트 앵커 D2 빨강을 스스로 알림.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: "preprep 레인 R1~R5·P1·P3·--baseline 이식 + known-pair + 실물 재현(워크트리 패치)"
+  count: 1
+  context: 덱 세션 인계분 ⓐⓑ. 중간에 지시 2회 갱신(R3 실물 쌍 · P1 이름 고유/3팔).
+  outcome: accepted
+  evidence: >-
+    scratchpad/preprep_lanes.patch → #628 커밋 2/3 — 24/24 · 실물 재현 6(40p·41p·38p·112p·36p ×2). 342,923 tokens · 57 tool_uses. 거버너 수리: 회사명 토큰(HIGH)·절대경로→env.
+  note: 이진 픽스처가 patch 로 안 실려 워크트리에서 직접 복사.
+- date: 2026-09-04
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-28
+  task: "preprep SKILL.md 방법론 7 · 정직한 한계 6 · 못 하는 것 표 · Done When · 체크리스트 C6(워크트리 패치)"
+  count: 1
+  context: 덱 세션 인계분 ⓔⓕ+§7. 레인 표는 손대지 말 것.
+  outcome: accepted
+  evidence: >-
+    scratchpad/preprep_docs.patch(128줄) → #628 커밋 3/3 — 새 레인을 이름(문자 코드 아님)으로 참조해 병렬 편집 충돌 회피. 194,195 tokens · 20 tool_uses.
+  note: 프론트매터 description 무변경(트리거 프로브 재실행 불필요).
+- date: 2026-09-05
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-29 (Fable 5.1 max 거버너, 오전)
+  task: "live-eval 12/12 FAILED-TO-RUN 수리 — timeout(1) 해석+bash 폴백 · fail-fast · reason 칼럼 · plist PATH · 레인(워크트리 패치)"
+  count: 1
+  context: 커밋 금지 워크트리. 원인은 거버너가 실측해 넘김(launchd PATH 에 timeout 없음). 중간 지시 1(FD_MODEL 템플릿 줄).
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/dispatch/2026-09-05_live-eval-timeout/patch.diff → 브랜치 fix/live-eval-timeout-preflight — 레인 31/31·28/28(fail-before 2+5 빨강) · env -i launchd PATH 라이브 1회 = bash-fallback 캡처 35B(FAILED-TO-RUN 아님). 388,607 tokens · 107 tool_uses.
+  note: 잔여 «USER 미설정 시 미인증」은 실 launchd 컨트롤(forge-wiki 위성 10:04 성공)이 반증 — env -i 가 launchd 보다 엄격. 거버너 추가 수리 1(워치독 orphan sleep pkill -P).
+- date: 2026-09-05
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-29
+  task: "다이제스트 variadic --allowedTools 인자 순서 수리 + 호출부 전수 감사 + 레인 ⑤(워크트리 패치)"
+  count: 1
+  context: 커밋 금지 워크트리. known-pair(haiku)는 거버너가 먼저 실측해 넘김.
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/dispatch/2026-09-05_digest-argorder/patch.diff → PR #631 — 레인 ⑤ 패치 전 ❌ → ✅ 5/5 · 호출부 9행 감사(취약 1). 272,961 tokens · 51 tool_uses.
+  note: REPORT.md Write 가 도구 정책으로 거부돼 회신 원문을 거버너가 옮김. 이후 에이전트에는 «Bash cat > 로 써라」 로 지시 변경.
+- date: 2026-09-05
+  agent: general-purpose (opus)
+  invoker: forge-harness-29
+  task: "qasp 2막 evidence 평문 입력값 두 채널 마스킹(__REDACTED__) — 구현·19 레인·커밋·푸시(PR 미개설)"
+  count: 1
+  context: qasp-dev 직접(필드 하네스, 자기 훅). 설계는 거버너가 확정해 넘김. 후속 지시 1(빈 값 비마스킹 한 줄).
+  outcome: accepted
+  evidence: >-
+    qasp-dev 3e76ff7 (fix/act2-evidence-secret-redaction) — tests/act2 1425→1444 · 전체 3714/0 · 되돌림 프로브 채널별 5/6 레인 · 셋째 채널(tc_update_suggestion) 발견·수리 · aria_snapshot 평문 실측 → 이슈 #253. 291,919 tokens · 82 tool_uses.
+  note: 설계가 «~14곳」이라 적은 것을 18곳으로 실측 정정. 보고서 tracks/qasp/secret_redaction_2026-09-05.md.
+- date: 2026-09-05
+  agent: sidecar codex exec (gpt-5.5, cross-family)
+  invoker: forge-harness-29
+  task: "qasp #254 비밀값 마스킹 diff 적대 리뷰 — 정규식 오탐·부분문자열 치환·to_dict/evidence_label 조건 동치·__RUNTIME__·테스트 공백"
+  count: 1
+  context: 페이로드 residency 스캔 TAINTED(2)→strip→CLEAN(3) 후 stdin 으로 diff 만 전달(파일 도구 없음). sidecar_wait 900s.
+  outcome: accepted
+  evidence: >-
+    scratchpad/codex_qasp/{prompt.txt,out.txt,wait.log} — SIDECAR_VERDICT 대기. 결과는 소스 그라운딩 후 PR #254 에 반영/반박.
+  note: ③ 기록면 토큰 첫 실물 = residency=CLEAN(files=3, stripped=2).
+- date: 2026-09-05
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-29
+  task: "⑫-③ residency 토큰 — pre-commit 검사 + 레인 r1~r13 + auto-decorrelation Step 4.5 + 출하(워크트리 패치)"
+  count: 1
+  context: 커밋 금지 워크트리. 중간 지시 1(strip 목록 명시). 스캐너가 워크트리에선 rc=10(override 패턴 부재) — 정직 기록.
+  outcome: accepted
+  evidence: >-
+    tracks/_meta/dispatch/2026-09-05_residency-token/patch.diff → feat/residency-marker-token — 55/55 · 되돌림 7 레인 · 픽스처 파일명 날짜 순서 결함 자력 적발. 363,065 tokens · 114 tool_uses.
+  note: 접두 grep 파싱이 codex 에서 fail-open 2건으로 지목돼 거버너가 presence-first 로 수리(r14~r17).
+- date: 2026-09-05
+  agent: sidecar codex exec (gpt-5.5, cross-family)
+  invoker: forge-harness-29
+  task: "⑫-③ 훅 변경 diff 적대 리뷰(6 파일, residency CLEAN(files=6))"
+  count: 1
+  context: load-bearing(pre-commit). diff 만 stdin. sidecar_wait 900s.
+  outcome: accepted
+  evidence: >-
+    scratchpad/codex_residency/out.txt — 8 findings: S 2 진짜(수리, 레인 r14/r15 fail-before 재현) · 4 잔여(과차단/fail-closed 방향) · B 2 레인. 27,937 tokens.
+  note: 거짓 양성 0 — 이번엔 페이로드를 안 잘랐다(strip=0).
+- date: 2026-09-05
+  agent: general-purpose (sonnet)
+  invoker: forge-harness-29
+  task: "⑫-② push-zone — pre-push 블록 + 마감 ①-f + 출하 + 레인 Z0~Z6(워크트리 패치)"
+  count: 1
+  context: 커밋 금지·stash 금지 워크트리. 실 git dispatch 로 첫 실사용.
+  outcome: partial
+  evidence: >-
+    tracks/_meta/dispatch/2026-09-05_push-zone/patch.diff → feat/push-zone-prepush 9e764c7 — 레인 13/13(fail-before 3/13) · 실 push 확인. 475,750 tokens · 174 tool_uses.
+  note: 🟥 `git stash` 1회 사용(지시 위반, 자기 신고, 영향 0) · ①-f 가 owner 이름을 그대로 출력(브리프의 «출력 그대로」 탓 — 거버너가 카운트만으로 수리) · URL 3형태만 수용(codex 4 형태 fail-open 지목).
+- date: 2026-09-05
+  agent: sidecar codex exec (gpt-5.5, cross-family)
+  invoker: forge-harness-29
+  task: "⑫-② pre-push push-zone diff 적대 리뷰(6 파일, residency CLEAN(files=6))"
+  count: 1
+  context: load-bearing(pre-push). diff 만 stdin. sidecar_wait 900s.
+  outcome: accepted
+  evidence: >-
+    scratchpad/codex_pushzone/out.txt — 11 findings: S 6·A 3 수리(레인 Z7~Z14 fail-before 10 적색 → 23/23) · B 1 잔여 · B 1 레인. 22,138 tokens.
+  note: 두 번째 codex 라운드 — 거짓 양성 0(strip=0).
+
+- date: 2026-09-05
+  agent: claude (opus, worktree fh-wt-outbound-hook)
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "⑫-① outbound query guard → PreToolUse(WebSearch|WebFetch) hook — two-layer degrade (override=deny · defaults=advisory), 76 lanes, 15 revert probes, clean-clone 2-arm run"
+  context_card: yes
+  outcome: accepted
+  evidence: "tracks/_meta/dispatch/2026-09-05_outbound-hook/REPORT.md — patch applied on feat/outbound-query-hook: 76/76 · 11/11 · degrade rc=0 · pkgcov PASS; agent self-caught .gitignore omission (clean-clone VOID) and a padding-timeout fail-open (10KB query = 49.7s in bash 3.2 `${var//[!\t]/}`); one false claim in report (main moved to 7530c9c — that is the #632 branch commit)"
+  tokens_subagent: 397555
+  notes: "governor follow-up: codex cross-family review dispatched on the payload; report claim about main position refuted by `git log -1 main`"
+
+- date: 2026-09-05
+  agent: codex exec gpt-5.5 (cross-family sidecar via scripts/sidecar_wait.sh)
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "adversarial review of the outbound_query_hook PreToolUse deny hook (9-file diff + author report, residency CLEAN)"
+  context_card: n/a (prompt file)
+  outcome: accepted
+  evidence: "11 findings: 6 S real (string tool_input, depth/count truncation x2, key-name log leak, severity-label leak, tr-dependent emptiness), 2 A real (N/A writes into stranger tree, lane timeout 15 vs snippet 5), 3 refuted (applicability path test, install path contract = same as five guards but noted, dedup assumption — measured: both-layer token denies with 2 hits). 8 fixes + 21 lanes; pre-fix hook goes 14 red on the new lanes"
+  tokens_subagent: n/a
+  notes: "first dispatch died in 5s: payload dir is not a git repo → needs --skip-git-repo-check; sidecar_wait typed verdict (exit=1 bytes=115) prevented reading it as a review"
+
+- date: 2026-09-05
+  agent: claude (opus, worktree fh-wt-utterance-intake)
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "engine reinforcement — utterance→record landing channel: transcript_utterances.py (shared with seal), utterance_intake.sh, session_close_check ①-f advisory, lanes, pre-registered floor-tier sim ARM/CTRL reps=3"
+  context_card: yes
+  outcome: accepted
+  evidence: "tracks/_meta/dispatch/2026-09-05_utterance-intake/REPORT.md — patch applied on feat/utterance-intake-close-check: lanes 51/51, seal byte-identical (L12 revert control), sim prereg sealed 9s before first arm; ARM 0/3 CTRL 0/3 with attribution withheld (3/3 chose memory/ outside the repo, runner isolation blocked it); first real use found 2 real unlanded operator utterances; mission-card error caught by lane L13 (①-e already taken → ①-f)"
+  tokens_subagent: 427744
+  notes: "5 design deviations all with measured grounds (min-chars 20, --strip-system-markers, record-derived positive control, private-token masking, ①-f label); 7 named residuals"
+
+- date: 2026-09-05
+  agent: codex exec gpt-5.5 (cross-family sidecar via scripts/sidecar_wait.sh)
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "adversarial review of the utterance-intake close-check patch (8-file diff + author report)"
+  context_card: n/a (prompt file)
+  outcome: accepted
+  evidence: "11 findings: 8 fixed (OR-probe false LANDED → per-token rows + majority; UNPROBEABLE silent rc=0; isMeta/isSidechain records; masking beyond $HOME; no-control-material message; L18c exit code; L13b absent-script branch; stale ①-e comment), 3 refuted (tool_result+text mixed 0/379 live; Hangul n-gram = judgment machinery; backtick comment carrier = pre-existing idiom). Pre-fix scripts go 15 red on the new lanes; fixed 72/72"
+  tokens_subagent: n/a
+  notes: "dispatched with --skip-git-repo-check (payload dir is not a git repo — recorded in memory project_codex_headless)"
+
+- date: 2026-09-05
+  agent: general-purpose (opus) — conference/industry insight scan
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "frontier absorption scan of 2025–2026 conference and industry material on AI quality gates / testing / red teaming, bookshelf-first (standards crosswalk) → 15 insights with FH/qasp delta, M/S/R, channel-vs-judgment tag; answer the operator's demand question as observable conditions; name the world's terms for gate intensity gears"
+  context_card: yes
+  outcome: accepted
+  evidence: "tracks/_meta/frontier_digest_2026-09-05_conference-insights.md (230 lines, 15 insights, sources marked opened/unopened, residency CLEAN with known-positive control). Top 3: gate intensity as a dial (Meta 4-level gating, arXiv:2410.06351), task validity / rule-blind arm (extension of ablation_calibrate arm B), unattended deployment missing from the irreversible-surface classes. Demand answer: 0/4 observable conditions today (one — STANDARDS_ALIGNMENT — lands with the open docs PR)"
+  tokens_subagent: 252151
+  notes: "author bias named in the file (search aimed at «deeper gates»); unopened sources listed by name; no commits"
+
+- date: 2026-09-05
+  agent: codex exec gpt-5.5 (cross-family sidecar via scripts/sidecar_wait.sh) — round 2 on the `oracle:` marker leg
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "adversarial review of validate_oracle_leg (bash 3.2/5), its 23-lane suite, wiring lane W6 and five doc surfaces — fail-open, over-block, near-miss claim, lane blind spots, portability"
+  context_card: n/a (prompt file, payload dir = scratchpad, residency CLEAN)
+  outcome: accepted
+  evidence: "NOT-CONVERGED: 8 findings (S2: enum member accepted as a mere prefix `known-pair2`; near-miss scan skipped when an exact `oracle:` line exists, so `Oracle: …` correction beside it was silent · A4 · B2). 6 fixed in v2 (always-on near-miss scan with one rule, delimiter-after-kind check, lanes o21–o34 with diagnostic needles, helper-absent lane o31, doc sync), 1 accepted as named residual (`<...>` literal in grounds), 1 doc-list unification. Self-detected 0/8. Mutants after fix: enum · delimiter · near-miss · word-count all turn the expected lanes red"
+  tokens_subagent: 112815
+  notes: "lesson saved to memory feedback_enum_prefix_and_shadowed_nearmiss"
+
+- date: 2026-09-05
+  agent: codex exec gpt-5.5 (cross-family sidecar) — round 3 on the fixed `oracle:` leg
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "verify the round-2 fixes adversarially (new failing inputs, over-blocking introduced by the always-on near-miss scan, surviving mutants, portability)"
+  context_card: n/a (prompt file, residency CLEAN)
+  outcome: accepted
+  evidence: "CONVERGED (S0 · A3 · B3). A: line starting with U+3000 under LC_ALL=C treated as absent · embedded CR hides a correction inside an excluded exact line · `oracle*` key namespace is reserved by the broad rule — all three recorded as named residuals / intended cost in the spec and AGENTS.md (first two are marker-wide, not leg-specific). B: PASS lane for `.` delimiter · BLOCK lane for an unlisted `oracle.evidence:` key · doc example lists marked non-exhaustive"
+  tokens_subagent: 37082
+  notes: "two rounds, R2→R3 delta was the fix set; no round 4 needed"
+
+- date: 2026-09-05
+  agent: Orca worktree session `qasp-map-archify` (Claude, Fable 5.1, effort medium — repo qasp-dev, operator-requested split)
+  invoked_by: FH governor session (Fable 5.1 max) — mission file + SendMessage return address
+  task: "leader-facing QASP whole-map in five layers (PAR · modules with real paths · trust/evidence · org-context boundary · web-tool vs CLI split) rendered with archify (local renderer, no LLM calls), plus Sister Asset three-part note; branch push only, no PR"
+  context_card: yes (MISSION.md with recipient, return address, output paths, prohibitions)
+  outcome: accepted
+  evidence: "branch chrono-meta/qasp-map-archify@5ebdaae — docs/map/QASP_MAP.md + 3 archify JSON/HTML (validate 9/9, visual-check 3/3), tests/test_map_paths_exist.py (20/20 real paths), ghe_paths_lib protected-path entry, README row; pytest 3766 pass / 14 env-dependent fail (0 code changes); fh-meta:beginner blind cold-read folded in; archify outbound check = update-check URL only (disabled). Three open questions escalated to the operator (web-tool first slice · docs/map protected-path vs company audience · PyMuPDF stdout breaking p1_schedule JSON)"
+  tokens_subagent: n/a (Orca session; it reported one beginner subagent at 40451)
+  notes: "phase 2 done: TC `oracle_type` on feat/tc-oracle-type — codex R1 NOT-CONVERGED S2·A3·B2 (parser dropped near-miss keys before the gate) → R2 NOT-CONVERGED S2 (pandas NA tokens folded to missing · non-string reason stringified) → R3 CONVERGED (A1 duplicate-header mangle, B1 whitespace header) → all fixed, 74/74 lanes, merged qasp-dev #261 (6c6f008) with the private companion store manifest. Map merged as #260 (ec418a6). Orca card completed"
+
+- date: 2026-09-05
+  agent: Orca worktree session `clawd-fh-gate` (Claude, Fable 5.1, effort medium — repo clawd-on-desk fork, expedition ⓑ-only)
+  invoked_by: FH governor session (Fable 5.1 max) — mission file + SendMessage return address
+  task: "map the auto-approve mechanism, design a three-layer gate (L1 irreversible hold · L2 output gate · L3 honest notice), prototype L1 with fail-before evidence; then a maintainer-standpoint review (static codex persona + dynamic isolated integration test); no push, no PR"
+  context_card: yes (MISSION.md; expedition CARD with 과녁 line, wall-clock, token slots)
+  outcome: accepted
+  evidence: "mechanism = app HTTP PermissionRequest hook layer (src/permission-automation-policy.js:387 → src/permission.js:2246); L1 prototype feat/auto-approve-irreversible-guard 56d64f3d→99f262a1, npm test 9423/0 fail, fail-before 20/24 red → green; maintainer-persona codex review S0/A2/B2 all fixed (shell-alias tools were outside the badge allowlist → contract test 13 rows); integration test 8/8 (fake HTTP → real server.js → real permission.js, unattended + force-push → DEFER observed); Electron live run correctly REFUSED (single-instance lock bound to userData, would collide with the production app). Operator's «FH gate 가 지켜줍니다» wording rejected as over-promise; honest 7-language notice instead. Wall-clock 14:40→16:40 KST; sidecar 1"
+  tokens_subagent: n/a (Orca session; sidecar codex ×1 unreported)
+  notes: "phase 2: maintainer-persona codex review S0/A2/B2 all fixed (99f262a1) + isolated integration test 8/8; PR_DRAFT.md in contributor voice (#888 style, no internal vocabulary). Live run on a second machine (operator's MacBook Air, Claude session over Remote Control, mixed human/AI driving): 10 rows as expected, 0 false positives, known-pair controls, reps=1 — the private companion store handoff/LIVE_TEST_RESULT_2026-09-05.md; three runbook defects found there and folded back. Outward PR submission = operator decision ㉓"
+
+- date: 2026-09-05
+  agent: fh-meta:hub-persona-auditor (opus) — pre-publication audit of an external-facing issue draft
+  invoked_by: FH governor session (Fable 5.1 max)
+  task: "audit ISSUE_DRAFT_external-gate.md (proposal to an external Electron app repo: optional external review command, downgrade-only) as maintainer / unattended power user / first-time contributor (+ a security reader it added), 4-axis per section, 3-tier revisions; judge length; list sentences a maintainer would quote back as over-claims; how to name our tool once without turning it into a pitch"
+  context_card: yes
+  outcome: accepted
+  evidence: "SHIP_AFTER_MANDATORY — mandatory 3 (complete the exit-code table incl. 124/126/127/128+n → ask; name the tool and drop 'exactly'; say what happens to an unanswered card), strong 6 (move the bypass-mode limitation to the opening, plain vocabulary, we→I, 7 over-claims softened, sessionId for sequence state, blast radius, 0/non-zero minimum), recommended 6 — all applied; final issue posted as rullerzhou-afk/clawd-on-desk#995 (816 words)"
+  tokens_subagent: 116852
+  notes: "auditor could not read the target repo's issue list (length verdict reasoned, not calibrated) — stated up front; one reader persona (security) added by the auditor itself"
+
+- date: 2026-09-05
+  agent: codex exec (gpt-5.5, read-only) — cross-family adversarial review, round 2, of the sync-to-be destination guard
+  invoked_by: FH governor session (Fable 5.1 max) — prompt via stdin (payload > ARG_MAX), residency-scanned first
+  task: "re-attack the R2 patch (root identity · physical paths · atomic --init · COUNT anchoring · timeout): any path where a wrong destination still receives a write, any write before refusal, COUNT-line spoofing, --no-git existence, timeout portability, bash 3.2/BSD"
+  context_card: yes (R1 findings + author's disposition per finding, R2 diff, post-patch lines 1-200, sync-from-be arg parser/COUNT site)
+  outcome: accepted
+  evidence: "NOT-CONVERGED — 7 findings: S1 guard validates _be_phys once but writes through mutable $BE (symlink retarget TOCTOU) · S2 default BE derived from the LOGICAL $FH · A3 old git's --show-superproject-working-tree swallowed · A4 independent nested repo inside a foreign tree passes · B5 COUNT line spoofable by a newline-bearing file name · B6 mid-run 'not a git repo' belt exits 0 after writes · B7 lane gaps. Disposition: S1/S2/A3/B6 fixed, A4 declined with grounds (lane B8m pins it), B5 named residual, B7 → lanes B8k~B8o; mutants m1~m4 each red only on the predicted lane. 59,799 tokens, 265s"
+  tokens_subagent: 59799
+  notes: "A3's direction was wrong in detail — measured: `git rev-parse --<unknown option>` ECHOES the option with rc=0 (git 2.50), so the old code refused with a nonsense reason rather than passing; both variants now fail closed with the real reason. The finding was still right that the branch was unexamined"
+
+- date: 2026-09-05
+  agent: codex exec (gpt-5.5, read-only) — cross-family adversarial review, round 3, same subject
+  invoked_by: FH governor session (Fable 5.1 max) — stdin payload (R3-only deltas + full post-R3 script + lane helpers), username scrubbed before send (2 hits → 0)
+  task: "verify each R2 closure against the code; attack the pin's coverage, the lock-wait injection lanes for flakiness/wrong-reason passes, the git shim, the A4 decline, new set -e interactions, bash 3.2/BSD"
+  context_card: yes
+  outcome: accepted
+  evidence: "NOT-CONVERGED — S1 the mid-run belt checked only --is-inside-work-tree, so a nested independent repo (accepted by A4) that loses its .git is still 'inside' the ENCLOSING repo and git add/commit/push would land the private half there (real hole, opened by the A4 decision) · A2 `_fh_phys0=$(_phys \"$FH\")` under set -e kills the script silently (rc=1) for a nonexistent HUB_DIR before the rc=10 hub refusal · B3 lane for the nested-repo mid-run case. All three taken into R4 (belt re-runs _be_root_ok; `|| true`; lanes B8o2/B8p; mutants m5/m6). 54,009 tokens, 190s"
+  tokens_subagent: 54009
+  notes: "the S1 here is a direct consequence of declining A4 in R3 — accepting the nested-repo layout created a belt that could pass for the wrong repo; the fix keeps the decision and tightens the belt to the same ROOT check the guard uses"
+
+- date: 2026-09-05
+  agent: codex exec (gpt-5.5, read-only) — cross-family adversarial review, round 4, same subject (sync-to-be destination guard)
+  invoked_by: FH governor session (Fable 5.1 max) — stdin payload (R4-only deltas + full post-R4 script), username scrubbed
+  task: "verify the R4 closures (belt re-runs the ROOT check; set -e safe derivation; lanes B8o2/B8p); attack the belt's re-check, the lock-wait lanes, bash 3.2 `cd -P \"\"` semantics, portability"
+  context_card: yes
+  outcome: accepted
+  evidence: "NOT-CONVERGED — S1 `.git` swapped mid-run for a gitfile pointing at another repo whose toplevel is still $BE passes the path-only re-check · S2 repo-selecting GIT_* env (a git hook exports GIT_DIR) makes a plain directory pass the root check · B3 lane for the swap. Also settled two of the author's own questions with measurements: bash 3.2.57 and 5.3.15 both FAIL `cd -P \"\"` (no no-op hazard), and A2's `|| true` reaches rc=10. Taken into R5: `unset GIT_*` first line + git identity pin (absolute-git-dir · common-dir) compared at the belt + threat model in the header + lanes B8q/B8r + mutants m7/m8 (m8 measured: without the unset the run mirrors into the plain dir and exits 4 from the hub-equality belt — wrong reason, after the write). 53,241 tokens"
+  tokens_subagent: 53241
+  notes: "S2 is the realistic one (hook environments); S1 needs an actor rewriting the store mid-run — accepted as a cheap non-over-blocking belt, and the header now states the threat model so later rounds grade against it instead of escalating the adversary each round"
+
+- date: 2026-09-05
+  agent: codex exec (gpt-5.5, read-only) — cross-family adversarial review, round 5, same subject (sync-to-be destination guard)
+  invoked_by: FH governor session (Fable 5.1 max) — stdin payload (R5-only deltas + full post-R5 script), username scrubbed; prompt now carries the THREAT MODEL so severity is graded against it
+  task: "verify R5 closures (GIT_* unset; git identity pin compared at the belt; lanes B8q/B8r; mutants m7/m8); attack the unset list's completeness, --git-common-dir comparability (relative vs linked worktree vs separate-git-dir), legitimate concurrent ops that change identity mid-run, bash 3.2 quoting, B8q's symlinked-tmp comparison"
+  context_card: yes
+  outcome: accepted
+  evidence: "CONVERGED — S0 · A0 · B1 (lane race: B8q waits for the 'destination:' line, which was printed BEFORE the identity pin, so a fast swap could be recorded as the baseline). Explicitly cleared: unset list covers repo selection (GIT_PREFIX does not redirect; CEILING/DISCOVERY vars only push toward fail-closed); common-dir comparison is same-cwd on both reads and does not reject legitimate separate-git-dir stores. B1 fixed as prescribed (pins moved before the destination line; log prints BE_REQUESTED → BE), lanes re-run + mutant m7 re-run; no R6 dispatched — the fix is the reviewer's own one-liner and the re-run is the verification. 67,325 tokens"
+  tokens_subagent: 67325
+  notes: "5 rounds total (R1 worktree draft → R5): S findings 1·2·1·2·0 — the count stopped falling until the threat model was stated in the header; stating it is what let the reviewer grade the mid-run-adversary class as B instead of escalating each round"
+
+- date: 2026-09-06
+  agent: Workflow `archify-contract-layer-reconstruction` (resume of wf_240055bb after the session limit) — 4 live agents (build:b-nodeps · verify:a-ajv · verify:c-table · judge) + 3 cached (seal-oracle · build:a-ajv · build:c-table); persona agents fh-meta:beginner ×2 reused from the map worktree
+  invoked_by: FH governor session (Fable 5.1 medium, overnight autonomous drive — operator «야간 자율주행 돌려줘 … qasp fh 모두 잘 부탁한다»)
+  task: "reconstruct archify's validate contract layer from the skeleton + verbatim schemas only (no source copy), three independent builds, verify each against the sealed 40-arm oracle (rc + code-set parity, copy-scan, fail-before), judge and write RECON_REPORT.md"
+  context_card: yes (workflow prompt carried the no-copy rule, sealed-oracle paths, threat model)
+  outcome: accepted
+  evidence: "parity b-nodeps 40/40 (zero deps, 1,068 LOC) · a-ajv 38/40 · c-table 38/40 (single fixture: throw-first vs collect-all reference policy) · copy-scan 0 hits on all three (planted-file control 73 hits) · fail-before 3/3 · 2/3 (fixture shadowing) · 3/3 · judge re-ran all 120 rows and reproduced the verifier numbers · reps=1 per arm. Tokens: 1,770,716 (resume) + 1,008,983 (pre-limit partial) · wall 30.6 + 21.7 min · builds preserved under expedition/recon-build/"
+  tokens_subagent: 2779699
+  notes: "first measured number for the operator's «코드 층 역공학 재구성» capability question — contract layer only; geometry/router/SVG not attempted (named). Chamber archify-recon verdict left to the operator (HITL); governor proposes CURATED (net-new = the measurement protocol, not a product)"
+
+- date: 2026-09-06
+  agent: general-purpose ×4 (qasp 조사 · oracle_type 배선 · 웹 백그라운드 동치 조사 · 오라클 스텝 배선 · PlaywrightAdapter 생명주기) + codex sidecar ×2 (cross-family 감사 2라운드)
+  invoked_by: FH governor session (Opus 5, 운영자 «qasp는 더 갈 부분있다면 그것도 시켜줘» → «모두 ㄱㄱ» → «검토 잘하고 머지»)
+  task: "qasp 다음 슬라이스 조사 → oracle_type 생산자 배선 · 2막 오라클 스텝/back verb 배선 · PlaywrightAdapter background/resume 첫 실 능력 · 각 건 cross-family(codex) 감사"
+  context_card: yes (레포 절대경로 · 정본 선독 지시 · 커밋 규율 · push/PR 금지 · 되돌림 컨트롤 요구)
+  outcome: accepted
+  evidence: "PR #264·265·266·268 전부 머지(main=01dafcd) · CI 6/6 × 4. cross-family 1차 A급 4건(3 수리 · 1 **반증** — needs_clarification 은 disposition 게이트가 이미 막는다) · 2차 S1+A4(전부 수리). 거버너 독립 되돌림 재검 4건 전부 적색 확인. 전체 스위트 1 failed/3996 passed(시간의존 픽스처 1, 기존)"
+  tokens_subagent: 986241
+  notes: "🟥 계기 확정 하나가 부수 수확: 「기존 실패 13건」은 코드가 아니라 venv 아티팩트였다(uv 3.11 위에 3.13 접목, pip 이 코드가 안 읽는 곳으로). 깨끗한 클론·라이브 트리 양팔 모두 1 failed. 세션마다 13/12/14/1 로 갈리던 원인. 🟥 그리고 codex 가 못 본 형제 경로(not-mapped 분기)를 거버너가 찾았다 — 사이드카는 지목자이지 판정자가 아니라는 계보 그대로"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrono-meta/fh-gate",
-  "version": "3.0.0",
+  "version": "3.1.0",
   "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": {
@@ -107,6 +107,9 @@
     "docs/IDENTITIES.md",
     "docs/GATE_DAY.md",
     "docs/platform_sustainability.md",
+    "docs/model_tier_expectations.md",
+    "docs/USE_CASES.md",
+    "docs/STANDARDS_ALIGNMENT.md",
     "knowledge/shared/GLOSSARY.md",
     "knowledge/shared/patterns",
     "knowledge/shared/plugin-catalog",
@@ -129,10 +132,15 @@
     "scripts/test_marker_floor_lanes.sh",
     "scripts/test_marker_defense_lanes.sh",
     "scripts/test_marker_crossfamily_lanes.sh",
+    "scripts/residency_closure_scan.py",
+    "scripts/test_residency_closure_lanes.sh",
     "scripts/test_marker_standpoint_lanes.sh",
     "scripts/test_marker_thirdparty_lanes.sh",
     "scripts/test_marker_axes_run_lanes.sh",
     "scripts/test_marker_soul_check_lanes.sh",
+    "scripts/test_marker_affected_lanes.sh",
+    "scripts/test_marker_oracle_lanes.sh",
+    "scripts/test_gate_two_verdicts_lanes.sh",
     "scripts/test_heavy_classifier_lanes.sh",
     "scripts/residency_admission_check.sh",
     "scripts/test_residency_admission_lanes.sh",
@@ -167,6 +175,7 @@
     "scripts/test_leak_scan_control_lanes.sh",
     "templates/predelete_check.sh",
     "templates/PRE-PUBLISH-CHECKLIST.md",
+    "templates/RED_TEAM_REPORT.md",
     "scripts/degrade_direction_scan.sh",
     "scripts/degrade_probe_capability.sh",
     "scripts/psa_probe_capability.sh",
@@ -178,6 +187,8 @@
     "scripts/prepush_guard_check.sh",
     "templates/predelete_check.test.sh",
     "scripts/test_prepush_destructive_lanes.sh",
+    "scripts/push_zone_check.sh",
+    "scripts/test_push_zone_lanes.sh",
     "scripts/fixture_guard_lib.sh",
     "scripts/shell_name_boundary_scan.sh",
     "scripts/test_verdict_watermark_lanes.sh",
@@ -202,6 +213,7 @@
     "scripts/test_prepush_stdin_integrity.sh",
     "scripts/public_surface_scan_files.sh",
     ".claude/rules/.public-surface-patterns.defaults",
+    ".claude/rules/.residency-patterns.defaults",
     "plugins/fh-meta/.claude-plugin/plugin.json",
     "plugins/fh-meta/skills",
     "plugins/fh-meta/agents",
@@ -234,6 +246,10 @@
     "scripts/stale_clone_guard.sh",
     "scripts/proposal_hook.sh",
     "scripts/test_proposal_hook_lanes.sh",
+    "scripts/outbound_query_hook.sh",
+    "scripts/test_outbound_query_hook_lanes.sh",
+    "scripts/outbound_query_guard.sh",
+    "scripts/test_outbound_query_lanes.sh",
     "scripts/test_revert_probe_lanes.sh",
     "scripts/revert_probe.sh",
     "scripts/test_preprep_drift_anchor_lanes.sh",
@@ -282,8 +298,30 @@
     "scripts/test_preprep_adjacent_dup_lanes.sh",
     "scripts/test_preprep_promise_lanes.sh",
     "scripts/test_mapped_tracks_lanes.sh",
+    "scripts/map_postprocess.py",
+    "scripts/test_map_postprocess_lanes.sh",
     "!**/__pycache__",
     "!**/__pycache__/**",
-    "!**/*.pyc"
+    "!**/*.pyc",
+    "scripts/probe_live_eval.sh",
+    "scripts/probe_live_eval_lib.py",
+    "scripts/test_probe_live_eval_lanes.sh",
+    ".claude/regression/probes_live.yaml",
+    "scripts/com.forge-harness.live-eval.plist",
+    "scripts/utterance_intake.sh",
+    "scripts/transcript_utterances.py",
+    "scripts/test_utterance_intake_lanes.sh",
+    "scripts/worktree_reclaim.sh",
+    "scripts/test_worktree_reclaim_lanes.sh",
+    "plugins/fh-qp/.claude-plugin/plugin.json",
+    "plugins/fh-qp/README.md",
+    "plugins/fh-qp/qp_profile.example.yaml",
+    "plugins/fh-qp/skills",
+    "plugins/fh-qp/scripts",
+    "plugins/fh-qp/fixtures",
+    "scripts/test_fh_qp_lanes.sh",
+    "scripts/test_preprep_diagram_lanes.sh",
+    "scripts/test_preprep_slide_refs_lanes.sh",
+    "scripts/test_action_yml_lanes.sh"
   ]
 }

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

@@ -1,6 +1,6 @@
 {
   "name": "fh-commons",
-  "version": "3.0.0",
+  "version": "3.1.0",
   "engines": {
     "claudeCode": ">=1.0.0"
   },

package/plugins/fh-commons/skills/preprep/README.md

@@ -6,7 +6,7 @@
 
 ```bash
 mkdir -p ~/preprep && cd ~/preprep
-cp <이 디렉터리>/{preprep.py,interslide_deps.py,lane_progression.py,lane_adjacent_dup.py,lane_promise.py} .
+cp <이 디렉터리>/{preprep.py,interslide_deps.py,lane_progression.py,lane_adjacent_dup.py,lane_promise.py,lane_diagram.py,diagram_from_json.py} .
 cp -r <이 디렉터리>/ooxml .          # 덱(pptx)을 다룬다면. 아니면 생략
 cp <이 디렉터리>/surfaces.example.yaml   surfaces.yaml
 cp <이 디렉터리>/canon_terms.example.yaml  canon_terms.yaml
@@ -16,6 +16,9 @@ python3 preprep.py
 ```
 
 의존: `pyyaml` · `python-pptx` 둘뿐이다(나머지는 표준 라이브러리). FH 도, 네트워크도 필요 없다.
+도해를 JSON 에서 굽는 갈래(`diagram_from_json.py`)만 **선택 의존**이 둘 더 있다 — node ≥ 18 + 로컬 설치한
+archify(`npx -y skills add tt-a1i/archify --skill archify --agent claude-code --copy --yes`, MIT) · headless Chrome.
+없으면 그 스크립트만 exit 10 으로 죽고 레인 L12 는 영수증 없음을 UNMEASURED 로 낸다 — 나머지 레인은 그대로 돈다.
 
 ## 드리프트를 막는 법