@chrono-meta/fh-gate 1.4.89 → 1.4.91

package/knowledge/shared/rules/multi_session_close_protocol.md

@@ -0,0 +1,118 @@
+# Multi-Session Close Protocol — 같은 하네스에서 둘 이상이 돌 때의 마감
+
+> **왜 여기 있나.** 이 프로토콜은 2026-08-09 까지 **세션 카드에만** 있었다. 카드는 gitignored 이고
+> **세대마다 통째로 덮인다** — 그날 실제로 이월 8건이 그렇게 사라졌고, 하필 그중 하나가
+> *"⑤-c 발화 착지 검증"*(카드가 내용 착지를 안 본다는 걸 고치려던 항목)이었다. 규율을 담은 곳이
+> 규율이 막으려던 방식으로 유실됐다. **그래서 tracked 로 옮겼다.** 이건 승격이 아니라 이사다.
+>
+> **범위**: FH 자신의 다축 마감. `templates/` 로 전파하지 않는다 — 필드 하네스는 대개 단일
+> 세션이라 아직 필요가 없고, 전파는 그쪽에서 병렬이 실제로 돌기 시작할 때다.
+
+## 0. 전제 — 두 가지 형상을 구분한다
+
+```
+거버너 + 위임      한 세션이 카드를 쥐고 나머지는 그 세션이 띄운 에이전트
+                   → 카드 소유가 명확하고 last-writer 경쟁이 없다. **기본형**
+대등 peer          독립 세션 둘 이상이 각자 마감한다
+                   → 아래 전부가 필요해진다. **다른 레포일 때만 권장**
+```
+2026-08-09 의 2축이 안전했던 이유는 모델이 좋아서가 아니라 **두 축이 다른 레포**였기 때문이다
+(`forge-harness` / `qasp-dev`). 같은 워킹트리를 공유하는 대등 peer 는
+`[[feedback_shared_checkout_ops_touch_others_work]]` 를 정면으로 부른다.
+
+## 1. 순서 — 질문이 재대조보다 **앞이고 더 넓다**
+
+```
+① peer 가 자기 델타를 append          (자기 소유 파일에만. 아래 §4)
+② 핑
+③ ★ 접기 전에 「지금부터 마감이다, 더 있나」를 **묻는다**
+④ 접는다 (⑤ 원자 실행: 로그 → 카드)
+⑤ `gh pr list --merged` 재대조
+```
+
+**③이 ⑤로 대체되지 않는다.** 재대조는 *내가 아는 축*만 훑고, 질문은 *상대가 아는 축*을 가져온다.
+2026-08-09 실측: 그 질문이 델타 8건을 꺼냈고 **PR 이 없어 `gh pr list` 로 구조적으로 안 잡히는
+것들**이 다수였다(스펙/설계만 바뀐 것 · PR 이전 브랜치 커밋 · 외부 승인 대기 · 이월 누락).
+
+## 2. 세 가지를 각각 못 믿는다
+
+### 2-a. 보냈다 ≠ 닿았다
+크로스세션 메시지는 **수신자 사용자 승인 대기로 보류**될 수 있고, 그래도 발신 호출은 성공을
+반환한다. 같은 날 카드는 이미 「5곳 전부 통지」라고 적고 있었다.
+
+⚠️ **만료 통지는 계기끼리 어긋났다 — 한쪽으로 단정하지 않는다.** 보류 2건 중 하나에 대해
+「승인 없이 만료 — 전달되지 않음」 통지가 왔는데, **운영자 확인으로는 도달해 있었다.**
+어느 쪽이 맞는지 이 세션은 못 가른다. 그러니 **«만료 = 영구 미도달» 로 적지 마라**(초안이 그렇게
+적었다가 반증됐다). 확실한 것은 하나뿐이다 — **발신 성공은 도달의 증거가 아니고, 통지도 아니다.**
+도달을 확인하는 유일한 방법은 **답이 오는 것**이다.
+→ **보낸 수가 아니라 답 온 수를 적고, 미회신 수를 카드에 명시한다.** 기계화 불가(스크립트는 회신을
+못 본다) — `session_close_check.sh` ①-c 가 이 문장을 살리언스로 띄운다.
+
+### 2-a-2. 탐지는 넓게, **통지는 좁게** — 둘을 같은 목록으로 쓰지 마라
+2026-08-09 의 오발신은 탐지가 넓어서가 아니라 **탐지 목록을 그대로 수신자 목록으로 썼기** 때문이다
+(원래 알려야 할 상대는 1곳이었다). 두 목록은 목적이 다르다:
+```
+탐지  넓게  — 살아있는 것 전부. 여분이 «마감 미완» 이라는 신호다 (§2-b)
+통지  좁게  — **지금 돌아가는(working/busy) 세션에만.** 끝난 세션에 보내는 건 소음이고,
+              수신자 승인을 태우며, 상대 기록에 남지도 않는다
+```
+**운영자 지시(2026-08-09)**: 「지금 돌아가는 동일 하네스 세션에만 보내라」.
+
+### 2-b. 살아있다 ≠ 일하는 중이다 · 그리고 그 여분이 신호다
+소켓이 살아있으면 **프로세스**가 살아있는 것이지 세션이 활성인 게 아니다. **판별 기준을 정확히
+적는다**: 소켓 파일명의 PID 에 프로세스가 살아있고 그 cwd 가 이 하네스면 잡힌다 — 그게 전부다.
+
+그래서 **세 가지가 섞여 들어오고, 통지가 필요한 건 첫째뿐이다**(2026-08-09 실측):
+```
+working 세션      → 접기 전에 물어라
+놀고 있는 터미널   → 진짜 세션이지만 프롬프트에서 대기 중. 보통 건너뛴다
+남은 프로세스     → 마감이 안 끝난 것 (1회 관측, 손으로 죽여야 했다)
+예열 워커(bg-spare) → **세션이 아예 아니다.** args 로 기계 제외 — 안 걸러내면 순수 오탐
+```
+⚠️ **초판은 여분 전체를 «마감 미완» 이라고 적었다. 한 사례에서 과일반화한 것이고 틀렸다** —
+실제 다수는 예열 워커와 놀고 있는 터미널이었다. **셋을 가르는 건 fleet 뷰뿐이고, 이 스캔은 못 한다.**
+⚠️ **job 상태파일로 좁히지 마라 — 반대로 틀린다.** 같은 날 실측: 상태파일이 **돌고 있는 세션을
+`done` + `firstTerminalAt` 로 표시**했다. 소켓이 넓은 만큼 상태파일은 좁다. **교차 판독.**
+
+### 2-c. 닫혔다 ≠ 안 열렸다
+「한 번도 안 닫힘」과 「닫고 또 일함」은 **밖에서 동일하게 Completed** 로 보이는데 처방이 반대다
+(전자는 마저 닫기, 후자는 다시 접기). `session_close_check.sh` ①-d 가 마감 게이트 통과를 세서
+2회 이상이면 알린다. **자기신고다** — 패턴을 감지하지 증명하지 않는다.
+
+## 3. 세대 간 승계 — 동료-질문으로는 구조적으로 안 잡힌다
+
+§1의 질문은 **동료 세션**을 향한다. **직전 세대 카드 자신의 이월분**은 그 질문의 사정거리 밖이다.
+2026-08-09 실측: 그렇게 사라진 이월이 8건이었고 파일은 전부 살아 있었다 — 소실이 아니라
+**카드에서만 안 잡히던 상태**.
+
+→ **카드를 재작성할 때 이전 카드의 이월 블록을 grep 대조한다.** 새로 쓰는 게 아니라 **덜어낸 것을
+확인하는** 절차다.
+
+## 4. 쓰기 규율 — 공유 자원은 append 도 충돌한다
+
+```
+✅ 자기 소유 파일 · 자기 섹션에만 append → **즉시 커밋**
+❌ `git add -A` (남의 미커밋 append 를 삼킨다)
+❌ 남의 카드 직접 편집 — 내용을 넘기고 소유자가 싣는다
+❌ `reset --hard` — Destructive-Op 게이트 대상
+```
+**메모리 인덱스도 같은 자원이다.** 2026-08-09: 한쪽이 항목을 archive 로 내리는 동안 다른 쪽이
+hot 에 재색인해 **양 tier 동시 등재**가 됐다. 적용 직후 검사는 0이었고 **나중에 겹쳤다** — 즉
+한 번의 검사로는 못 잡는다.
+
+## 5. 마감 전 마지막 — 내용 착지를 따로 검증한다
+
+close check 는 **형식**만 본다(카드가 로그보다 새로운가, 필수 아티팩트가 있는가). **내용이
+착지했는지는 안 본다.** 2026-08-09 에 `CONSISTENT` 를 열세 번 내고도 발화 2건이 미기록이었고,
+`scripts/utterance_landing_check.sh`(컨트롤 동반 grep)가 처음 잡았다.
+
+**가장 잘 빠지는 것은 잊은 발화가 아니라 «행동으로 대응한 발화»다** — 답장에서 조사·판단까지
+하고 나면 처리된 느낌이 남고, 그 판단이 어디에도 안 적힌 채 세션이 끝난다.
+
+## 명명된 잔여
+
+- ①-c 의 peer 판별은 **cwd 매칭**이라 **다른 worktree 의 peer 를 못 잡는다**(Orca 류 격리 사용 시).
+- ①-d 스탬프와 §5 검사 결과는 **자기신고**다. 형식은 잡고 위조는 못 잡는다.
+- §1 의 질문은 **사람/세션이 해야 한다** — 기계 트리거가 없다. ①-c 는 «물어라» 를 띄울 뿐이다.
+- 이 문서 자체가 n=1(2026-08-09 2축 마감 1회) 위에 서 있다. **두 번째 사례가 나오면 다시 봐라** —
+  특히 §0 의 「거버너+위임이 기본형」은 그 한 번에서 나온 판단이다.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@chrono-meta/fh-gate",
-  "version": "1.4.89",
+  "version": "1.4.91",
   "description": "FH runtime adapters — run FH governance, skills, and agents via Claude or Codex with machine-parseable gates.",
   "license": "MIT",
   "keywords": [
@@ -90,10 +90,22 @@
     "scripts/substrate_jump_detector.sh",
     "scripts/tier_census_grep.sh",
     "scripts/test_marker_floor_lanes.sh",
+    "scripts/test_marker_crossfamily_lanes.sh",
+    "scripts/directional_diff_gate.sh",
+    "scripts/reviewer_capability_corpus.tsv",
+    "scripts/test_reviewer_capability_conformance.sh",
     "plugins/fh-meta/CHANGELOG.md",
     "knowledge/shared/learnings/subagent_invocations_log.yaml",
     "scripts/chamber_candidate_collect.sh",
+    "scripts/chamber_witness.sh",
+    "scripts/digest_landing_check.sh",
+    "scripts/relay_channel.sh",
+    "scripts/test_relay_channel_lanes.sh",
     "scripts/fh_session_load.sh",
+    "scripts/branch_claim.sh",
+    "scripts/test_branch_claim_lanes.sh",
+    "scripts/hook_source_lib.sh",
+    "scripts/test_hook_source_gate_lanes.sh",
     "templates/.claude/rules/mcp_tool_gating.md",
     ".claude/rules/fh_4axis_gate.md",
     "templates/degrade_direction_scan.sh",
@@ -107,6 +119,7 @@
     "scripts/prepush_guard_check.sh",
     "scripts/psa_scan_lib.sh",
     "scripts/session_close_check.sh",
+    "scripts/utterance_landing_check.sh",
     "scripts/test_dispatch_log_lanes.sh",
     "scripts/memory_link_check.py",
     "scripts/test_memory_link_check.sh",
@@ -154,6 +167,14 @@
     "templates/subagent-tally-hook.json",
     "scripts/test_session_close_chain_lanes.sh",
     "scripts/test_node_infra_delta_lanes.sh",
-    "scripts/test_wizard_snippet_merge_lanes.sh"
+    "scripts/test_wizard_snippet_merge_lanes.sh",
+    "scripts/compaction_probe.sh",
+    "scripts/judgment_circuit_lint.sh",
+    "scripts/novelty_claim_check.sh",
+    "templates/settings.Compaction.snippet.json",
+    "scripts/field_canon_preload.sh",
+    "scripts/test_field_canon_lanes.sh",
+    "templates/settings.FieldCanon.snippet.json",
+    ".claude/judgment_circuits.txt"
   ]
 }

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

@@ -1,6 +1,6 @@
 {
   "name": "fh-commons",
-  "version": "1.4.89",
+  "version": "1.4.91",
   "engines": {
     "claudeCode": ">=1.0.0"
   },

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

@@ -1,6 +1,6 @@
 {
   "name": "fh-meta",
-  "version": "1.4.89",
+  "version": "1.4.91",
   "engines": {
     "claudeCode": ">=1.0.0"
   },

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

@@ -197,13 +197,54 @@ mode**, with an honest *same-family* note but **no "reduced value / degraded" fr
 this floor. `unset` → ask-once at first load-bearing need (accept → proceed; decline → record + this
 branch). Only proceed to the discovery ladder below when consent is `accepted`.
 
-1. frontier cross-family CLI present → recruit it (decorrelated, at-floor) — best.
-2. only local 4090 present → canary pre-screen + in-session opus governor (canary, not full decorrelation).
-3. nothing present → in-session same-family + **honest below-floor/same-family note** (residual named).
-
-Env non-determinism (CLI presence varies) → **silent degrade, never hard-fail**. Distinguish this
-**unavailable-but-wanted** case (consent given, panel down → degrade-with-note; for a *load-bearing corp*
-surface, fail-closed per `local_pmh_context.md`) from the **declined** case above (chosen floor, first-class).
+Every rung **emits a typed `crossfamily:` value into the Axes 2–3 marker** — the rung is not
+done until its verdict is recorded. The value is a closed enum, validated by `pre-commit`
+(`validate_marker_floor`, fixtures in `scripts/test_marker_crossfamily_lanes.sh`); free prose is
+rejected at commit time.
+
+| # | Panel state | Emit | Ack |
+|---|---|---|---|
+| 1 | frontier cross-family CLI present → recruit it (decorrelated, at-floor) — best | `panel(<families>)` | — |
+| 2 | only local 4090 → canary pre-screen + in-session opus governor (canary, not full decorrelation) | `panel(<families>)` **only if** a capable non-Claude model actually reviewed; otherwise rung 3 | — |
+| 3 | **probed and nothing capable reachable** → in-session same-family | `DEGRADED_SINGLE_FAMILY` | **required** |
+| — | **capable panel reachable, not recruited** (a choice, not a constraint) | `DEGRADED_PANEL_UNUSED` | **required** |
+| — | consent `declined` (branch above — chosen floor, not a degrade) | `declined` | — |
+| — | change is not load-bearing — decorrelation not required | `single-family` | — |
+| — | **panel never probed** | `UNKNOWN` | **required** |
+
+**The three degrade values are the load-bearing split**: *could not* (`DEGRADED_SINGLE_FAMILY`) ·
+*did not* (`DEGRADED_PANEL_UNUSED`) · *did not look* (`UNKNOWN`). Free prose merges all three, and
+each merge hides a different thing — an unrun probe renders as a zero finding
+(`[[feedback_not_found_is_not_zero_family]]`), and an unused panel renders as an unavailable one.
+The sibling-harness signal that motivated this field said it in its own words about itself:
+*"못 한 것이 아니라 안 한 것이다."* All three require an ack naming *what was probed and what it
+returned*. **`DEGRADED_PANEL_UNUSED` was added by dogfooding**: the first real marker written under
+this lane found the panel reachable (codex · agy · gemini) and no cross-family review run — a state
+the original five-value enum could not express without lying
+(`[[feedback_adversarial_review_not_substitute_for_first_use]]`).
+
+**Panel membership is capability-gated, and ineligibility is tested FIRST.** A runtime `/models`
+list is a **serving** list, not a generative one. Measured in a sibling harness (2026-08-08): 6 of
+14 served entries could not review anything (embedding · reranker · OCR · safeguard), and because
+the embeddings' ids matched no family prefix **each counted as its own family — satisfying the
+family-diversity check while the panel was incapable**. Order is the invariant, not the list:
+`GLM-OCR` matches both `ocr` and `glm`, `Qwen3-Embedding` matches both `embed` and `qwen`, so an
+eligibility-first test lets them through silently. If nothing capable survives the filter, that is
+rung 3 — **not** a smaller panel.
+
+**Attribution: `panel(...)` records the families *requested*, not confirmed.** At least one
+procurement path (an internal corp-gateway CLI) returns **no model attribution in its response**, so "3/3 responded"
+does not establish which model contributed. Record request-fact and response-fact separately; do not
+launder the former into the latter (`[[feedback_instrument_cannot_discriminate_hypotheses]]`).
+
+Env non-determinism (CLI presence varies) → **never hard-fail the run** — but never *silent* either.
+The two are different properties, and the prose this replaces conflated them: the degrade proceeds
+(a commit is a reversible surface) **and** lands in the verdict with an ack. That is how this rung
+and `CLAUDE.md` §Field-Harness gate ("never a silent same-family pass") are simultaneously true.
+Distinguish the **unavailable-but-wanted** case (consent given, panel down → `DEGRADED_SINGLE_FAMILY`
++ ack; for a *load-bearing corp* surface, fail-closed per `local_pmh_context.md` — note that file is
+**pmh-repo-local and unreadable from an FH session**, so on this side the typed value + ack is the
+enforceable layer) from the **declined** case above (chosen floor, first-class, no degrade framing).
 
 ## Step 7 — Output
 
@@ -221,7 +262,14 @@ at N per run (no loop).
 
 - Mechanical sidecar discovery returns the available panel. *[mandatory-pass]*
 - Family-diversity recruitment picks ≥1 cross-family verifier when present (family by model probe). *[mandatory-pass]*
-- Degrade ladder applied: local-only → canary+note · none → same-family note. *[mandatory-pass]*
+- Degrade ladder applied AND its verdict emitted as a typed `crossfamily:` value in the Axes 2–3
+  marker (`panel(<families>)` | `declined` | `DEGRADED_SINGLE_FAMILY` | `DEGRADED_PANEL_UNUSED` |
+  `UNKNOWN`); the three degrade values carry substantive grounds on the same line. `single-family`
+  is NOT accepted on a load-bearing change — that block only runs there, so "decorrelation not
+  required" is a contradiction, and as a no-ack pass it was a free bypass of the lane.
+  *[mandatory-pass — enforced by pre-commit, fixtures `scripts/test_marker_crossfamily_lanes.sh`]*
+- Panel members are review-capable: embedding/reranker/OCR/safeguard classes are excluded **before**
+  family-diversity is counted, never after. *[mandatory-pass — same fixtures, ordering anchor c4/c5]*
 - Paid recruit was per-run spend-gated (or `paid_auto` set); free local tier may auto-fire. *[mandatory-pass]*
 - Governor (CC) retains terminal verdict; every accepted sidecar finding is source-grounded. *[judged, pair: judge-robustness / phantom-quench back-trace]*
 - Marker records panel + families only with a captured sidecar transcript. *[mandatory-pass]*

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

@@ -170,6 +170,39 @@ Execution priority (behavioral):
 
 > **Detail**: See `SKILL_detail.md §Step3-Execution` — action preview example, FH plugin auto-install bash + error handling table + success format, agent-composer mapping update format — read when executing Step 3.
 
+### Step 3-D. Dispatch Standing Request — the one moment this is cheap to settle
+
+FH/PMH treat isolated delegation as part of what the harness is, so **parallel dispatch is the
+intended default for anyone running them**. It is not default *by assertion*: some runtimes ship a
+line making unprompted dispatch conditional on a user request, and `CLAUDE.md §Agent Dispatch
+Operation` will not read that condition as met unless a standing request is on disk. **A session with
+no such record asks before every dispatch** — which is the friction this step exists to remove, once,
+here, instead of in every future session.
+
+So ask, and write the answer down. Three parts, all required — a record missing one is not a record,
+and a session may never supply the missing part itself:
+
+```
+① the user's own words, quoted   — ask them to say it, then quote what they said
+② a lease                        — THE USER picks the length. Do not suggest a default and
+                                   do not fill it in. Renewal needs a fresh utterance later
+③ a scope                        — which prohibition line this covers. Granting one never
+                                   grants another (agents ≠ workflows ≠ deep-research):
+                                   name each one the user actually intends
+```
+
+Write it into `CLAUDE.local.md` (gitignored). **Declined is a valid, recorded answer** — write that
+down too, so later sessions stop re-asking rather than reading silence as an unanswered question.
+
+**Any mechanical settings change belongs here as well**, under the same consent: if enabling the
+intended default needs a `settings.json` edit, propose it in this step with the diff shown, and apply
+it only on approval. Do not change settings silently on the strength of the doctrine alone — the
+doctrine says dispatch is the default posture, not that the wizard may edit a user's config unasked.
+
+⚠️ **The lease is honoured by reading, not by machinery.** `scripts/consent_registry_check.sh`
+enforces leases for *registered consent classes*, and this grant is not one of those — say so when
+recording it rather than implying an expiry alarm exists.
+
 ### Step 4. Acceleration Baseline Setup
 
 After executing approved items, install the automatic maintenance structure: zshrc hook (idempotent, preview-then-confirm) + sentinel initialization (per-project) + weekly-audit cadence (no cron — zshrc hook + session-start detection are the durable mechanisms).

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

@@ -526,31 +526,120 @@ fi
 # scripts/package_coverage_check.sh caught the omission — without it an npm-installed wizard hit
 # FileNotFoundError here and registered nothing while reporting success upstream).
 python3 - "$FH_DIR" <<'PY'
-import json, os, sys, collections
+import json, os, sys, glob, collections, re, shutil
 hub = sys.argv[1]
-snippet = os.path.join(hub, "templates", "settings.SessionStart.snippet.json")
 target = os.path.join(hub, ".claude", "settings.json")
-entry = json.load(open(snippet))["project_settings_json"]["hooks"]["SessionStart"]
+
+# DISCOVER every shipped snippet, do not name one. Hardcoding a single snippet path is why the
+# compaction hooks shipped unwired while their own README recited the shipping-is-not-wiring lesson
+# (high review, 2026-08-08 — and PreToolUse was already sitting in the same hole, so N=2).
+# A new templates/settings.*.snippet.json is picked up with zero edits here: generation, not detection.
+snippets = sorted(glob.glob(os.path.join(hub, "templates", "settings.*.snippet.json")))
+if not snippets:
+    # exit 1, NOT 0. The pre-rewrite code raised FileNotFoundError here, and that loud failure is what
+    # made a broken npm `files[]` detectable — the comment above records it happening for real once.
+    # The first rewrite replaced it with a quiet exit 0, i.e. it deleted the detector while keeping the
+    # bug (high re-review 2026-08-08 #1). Zero snippets means zero hooks registered; that is never a
+    # success, and every other zero-registration path below already exits 1.
+    print("NO SNIPPETS under templates/settings.*.snippet.json — zero hooks registered.")
+    print("  → package `files[]` may have dropped templates/, or this is not an FH hub. NOT a success.")
+    raise SystemExit(1)
+
 d = collections.OrderedDict()
 if os.path.exists(target):
     d = json.load(open(target), object_pairs_hook=collections.OrderedDict)
-    import shutil; shutil.copy2(target, target + ".prewizard")   # back up before rewriting
+    shutil.copy2(target, target + ".prewizard")          # back up before rewriting
 hooks = d.setdefault("hooks", collections.OrderedDict())
-# Merge at HOOK level, not group level. A group-level filter drops the whole group when a user's own
-# hook shares a group with the FH one — the common shape when someone hand-edits or appends to an
-# older wizard's output. (Cross-family review 2026-07-30 reproduced the loss: a group holding
-# [my_telemetry.sh, fh_session_load.sh] lost my_telemetry.sh entirely.)
-kept = []
-for g in hooks.get("SessionStart", []):
-    survivors = [h for h in g.get("hooks", []) if "fh_node_check" not in h.get("command", "")]
-    if survivors:
-        g = dict(g); g["hooks"] = survivors; kept.append(g)
-hooks["SessionStart"] = kept + entry
+
+def fh_keys(entry_groups):
+    """FH script basenames this snippet owns — the replace key is DERIVED from the snippet's own
+    commands, never hardcoded, so it cannot drift from what is actually being installed."""
+    # Key on the snippet's own PATH FORM (scripts/<name>.sh), not a bare basename. A bare basename
+    # made the survivor filter delete a user's own hook that merely shared a filename — e.g. their
+    # ~/tools/compaction_probe.sh — which is the exact loss the hook-level merge exists to prevent,
+    # reintroduced through key derivation (high re-review 2026-08-08 #2).
+    keys = set()
+    for g in entry_groups:
+        for h in g.get("hooks", []):
+            for m in re.findall(r'((?:[A-Za-z0-9_\-]+/)*scripts/[A-Za-z0-9_\-]+\.sh)', h.get("command", "")):
+                keys.add("scripts/" + m.rsplit("scripts/", 1)[1])
+    return keys
+
+registered, bad_schema, intended = [], [], {}
+for snip in snippets:
+    try:
+        blob = json.load(open(snip))
+    except Exception as e:
+        print("SKIP (unparsable):", os.path.basename(snip), e); bad_schema.append(snip); continue
+    proj = (blob.get("project_settings_json") or {}).get("hooks") or {}
+    if not proj:
+        # settings_local_json_MODE_D_ONLY entries carry private paths — never auto-written here.
+        print("SKIP (no project_settings_json):", os.path.basename(snip)); bad_schema.append(snip); continue
+    for event, entry in proj.items():
+        # Shape-validate before touching the user's file. A valid-JSON / invalid-SCHEMA snippet used
+        # to be written anyway and reported as success, then detonate on the NEXT run inside the
+        # survivor filter — far from the snippet that caused it (lanes BS-*, LT-*).
+        if not isinstance(entry, list) or not all(
+                isinstance(g, dict) and isinstance(g.get("hooks"), list) for g in entry):
+            print("SKIP (bad schema):", os.path.basename(snip), event); bad_schema.append(snip); continue
+        keys = fh_keys(entry)
+        if not keys:
+            print("SKIP (no FH script in commands):", os.path.basename(snip), event)
+            bad_schema.append(snip); continue
+        intended[event] = intended.get(event, set()) | keys
+        # Merge at HOOK level, not group level. A group-level filter drops the whole group when a
+        # user's own hook shares a group with the FH one (cross-family review 2026-07-30 reproduced
+        # the loss: [my_telemetry.sh, fh_session_load.sh] lost my_telemetry.sh entirely).
+        kept = []
+        for g in hooks.get(event, []):
+            survivors = [h for h in g.get("hooks", [])
+                         if not any(k in h.get("command", "") for k in keys)]
+            if survivors:
+                g = dict(g); g["hooks"] = survivors; kept.append(g)
+        hooks[event] = kept + entry
+        registered.append(f"{event}({','.join(sorted(keys)) or '?'})")
+
 os.makedirs(os.path.dirname(target), exist_ok=True)
 with open(target, "w") as fh:
     json.dump(d, fh, indent=2, ensure_ascii=False); fh.write("\n")
-print("node-check SessionStart hook registered ->", target)
+# RE-READ WHAT WE JUST WROTE. The success message used to be unconditional — it never checked the
+# serialized result, so a snippet that failed to register still printed "registered ->". Verify, then
+# claim (`[[feedback_gate_verification_must_execute]]`).
+written = json.load(open(target))
+missing = []
+for event, keys in intended.items():
+    blob = json.dumps(written.get("hooks", {}).get(event, []))
+    for k in sorted(keys):
+        if k not in blob:
+            missing.append(f"{event}:{k}")
+for r in registered: print("    hook registered ->", r)
+print("snippets scanned:", len(snippets), "· hook entries registered:", len(registered),
+      "· skipped:", len(bad_schema), "->", target)
+if missing:
+    print("FAILED to register:", ", ".join(missing)); raise SystemExit(1)
+# FAIL CLOSED on any skipped snippet. The good ones are already written (a malformed sibling must not
+# cost you the working hooks), but the RUN exits non-zero — an installer that skips a snippet and
+# reports success leaves hooks unwired, which is the exact defect this rewrite exists to close.
+# Lanes MG-N1..N4 caught this: the first draft skipped-and-exited-0 = fail-open.
+if bad_schema:
+    print("SKIPPED snippet(s):", ", ".join(sorted({os.path.basename(b) for b in bad_schema})))
+    print("  → non-zero exit on purpose: a skipped snippet is unregistered hooks, not a warning.")
+    raise SystemExit(1)
+if not registered:
+    print("no hook entries registered — nothing to claim"); raise SystemExit(1)
 PY
+# CONSUME the merge block's exit status. It raises SystemExit(1) on a skipped/failed registration,
+# and the surrounding bash has no `set -e` — so without this check the wizard sailed past a failed
+# registration and §Step5 still printed success, which is the exact "reports success while unwired"
+# shape the fail-closed direction was added to prevent (high re-review 2026-08-08 #6).
+# The failure is SURFACED, not swallowed: the operator sees which snippet did not register.
+_MERGE_RC=$?
+if [ "$_MERGE_RC" -ne 0 ]; then
+  echo "🟥 hook registration FAILED (exit $_MERGE_RC) — one or more snippets are UNREGISTERED."
+  echo "   Do not read the completion report below as 'hooks are wired'. Fix the snippet and re-run"
+  echo "   this step; the rest of the wizard continues so the remaining setup is not lost."
+  WIZARD_HOOK_REG_FAILED=1
+fi
 chmod +x "$FH_DIR/scripts/fh_node_check.sh" 2>/dev/null
 # VERIFY against a THROWAWAY state file (FH_NODE_STATE). Verifying against the real state would
 # consume the user's one-shot event report, so their actual first session goes quiet and the notice