@chrono-meta/fh-gate 1.4.95 → 1.4.97

package/plugins/fh-commons/skills/mcp-circuit-breaker/SKILL.md

@@ -182,13 +182,22 @@ Sending HALF-OPEN probe to {tool-name}...
 - If success: circuit → CLOSED, log updated
 - If fail: circuit remains OPEN, cooldown resets
 
-Reset log entry:
+Reset log entry — **append as a NEW list item (`- tool:`), never as bare indented keys**:
 ```yaml
+- tool: {tool-name}
   state: CLOSED
   reset_at: {ISO-8601}
   reset_method: probe_success | user_forced
 ```
 
+> **Why the leading `- tool:` is load-bearing.** The Step 3 log is a YAML *list*. Appending indented
+> keys with no `-` merges them into the **previous** item instead of adding one, and PyYAML resolves
+> duplicate keys silently by keeping the last value — so `state: OPEN` is overwritten by `CLOSED` and
+> the trip record disappears. Measured 2026-08-11: trip entry + bare-key reset entry parses to
+> **1 item** with `state: CLOSED` (the OPEN row is gone); the same pair written as two list items
+> parses to **2 items** with the trip history intact. The state history is what Step 6 reports on, so
+> a merged entry silently zeroes the failure record.
+
 ---
 
 ### Step 6. Report

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

@@ -1,6 +1,6 @@
 {
   "name": "fh-meta",
-  "version": "1.4.95",
+  "version": "1.4.97",
   "engines": {
     "claudeCode": ">=1.0.0"
   },

package/plugins/fh-meta/CHANGELOG.md

@@ -10,6 +10,42 @@ Format: [Keep a Changelog](https://keepachangelog.com/en/1.1.0/)
 
 ## Plugin Level
 
+### [1.4.96] — 2026-08-12
+
+**fix: 전수 재출하 캠페인 — 「돌았다고 보고하는데 안 도는」 계기들 · 선언이 무효였던 에이전트**
+
+스킬 40종 + 에이전트 8종 전수 적대검토 후 수리(PR #348 · #349). 지배 결함은 하나의 얼굴이었다 —
+**계기가 대상에 안 닿는데 출력은 무결**.
+
+- **`quench-challenger` 의 frontmatter 가 깨져 선언이 무효였다.** 비인용 멀티라인 `description`
+  안의 `user:`/`assistant:` 줄이 새 YAML 키로 파싱돼 그 아래 `tools: Read, Grep, Glob` 과
+  `model: opus`(HARD FLOOR) 를 **둘 다 무효화**했다. 읽기전용으로 선언된 적대 에이전트가 전체
+  도구로 돌고 있었다. `description` 을 한 줄 인용으로, 예시는 본문으로.
+- **`scripts/validate_yaml.sh`** — 스캔 대상에 **에이전트 8종 편입**(종전 SKILL.md 만). 양 surface
+  중 하나라도 0건이면 `INSTRUMENT ERROR`(exit 3) — 스캔 0을 통과로 렌더하지 않는다. **신규 출하**.
+- **`scripts/degrade_direction_scan.sh`** — **마크다운 ```bash 펜스 추출**. FH 가 실제로 실행하는
+  bash 는 대부분 SKILL.md 펜스 안에 사는데 스캐너가 `.py`/`.sh` 만 봤다. 그림자 파일이 **원본
+  줄번호를 보존**해 findings 가 `SKILL.md (```bash fence):202` 로 나온다. 펜스 없는 md 는 종전대로
+  `UNSCANNABLE`(미측정을 커버리지로 바꾸지 않는다).
+- **인용 무결성 2건 정정**(원문 직독) — arXiv 2605.00914 의 32.3pp 는 **다수결 oracle gap** 이지
+  「자기평가 시 성능저하」가 아니고(논문 결론은 *isolated self-correction prevails*), arXiv
+  2603.15255(SAGE)는 **co-evolve** 라 「Critic 격리」 근거가 될 수 없다. `harvest-loop` 의 같은
+  오귀속도 동시 수리.
+- **`salience-splitter`** — 4축 게이트가 이름으로 지목한 의무(«split 마다 목적지가 게이트 안인지
+  재질문») 를 본문에 배선 · 컷 판정 「머릿속으로」를 **레이어별 측정 2분기**로(상주=ablation
+  하네스 / SKILL.md=콜드스타트 sim) · 포인터↔헤더 대조를 실행 가능한 형태로(오탐 100% 였다) ·
+  orphan 스캔을 `^## §` → `^## ` 로(실물 헤더 139 중 97만 보던 30% 사각).
+- **죽은 명령·경로 정리** — `claude mcp search`(부재) · `codex list-agents`(부재) ·
+  `marketplace add` 인자 · `GoogleWebSearch`(존재하지 않는 도구명) · `frontier-digest` 저장 경로가
+  카덴스 글롭과 어긋나 **7일 카덴스에 영원히 안 잡히던** 것.
+- **거짓 PASS 계기들** — `harness-doctor` E7/E3/Step-11(`grep -c || echo 0` 일가가 음성 arm 을
+  통과로 렌더) · `install-doctor` 의 `except: pass`(깨진 MCP 설정이 「위험 없음」과 구별 불가) ·
+  `asset-placement-gate`(죽은 스캔과 진짜 무충돌이 둘 다 count=0) · `auto-decorrelation`(zsh 에서
+  다중 사이드카가 **무음 탈락** → single-family degrade).
+- **`AGENTS.md`** — 도구 표 누락 3종(beginner·main-player·expert) 보강 + frontmatter YAML 유효성
+  규율(비-CC 파서에서는 깨진 줄 아래 키가 전부 드롭된다). 문서↔파일 전수 대조 불일치 0.
+
+
 ### [1.4.86] — 2026-08-03
 
 **fix: 격리를 자칭하던 어블레이션 절차 + ambig 게이트 앵커 + 밀린 출하 자산 반영**

package/plugins/fh-meta/agents/beginner.md

@@ -89,7 +89,10 @@ Did I reach a successful first outcome? : YES / NO (blocked at [where])
 ## Integration Hooks
 
 - **sim-conductor Area A** — beginner is the canonical first-contact persona (A-1). Pairs with main-player (engaged use) and challenger (adversarial).
-- **marketplace-gate / install-wizard** — README & onboarding-path friendliness cold-read.
+- **marketplace-gate / install-wizard** *(claimed adjacency, not wired — measured 2026-08-11)* —
+  README & onboarding-path friendliness cold-read. `grep -rc beginner` → marketplace-gate **0**,
+  install-wizard 1 (a prose mention in SKILL_detail, not a dispatch). Control: sim-conductor lists
+  this agent in its persona table 4×. Treat as a *fit*, not an integration.
 - **hub-persona-auditor boundary** — *lens*, not artifact-exclusivity. Both may touch a README. `hub-persona-auditor` = multi-reader 4-axis pre-publication audit of external-facing drafts (briefing/card/guide/README as a publication). `beginner` = a single cold-read standpoint surfacing first-contact friction in any artifact (SKILL/README/prompt/config/code). For a publication-readiness verdict on an external draft, defer to hub-persona-auditor; for "can a first-timer actually get started," use beginner.
 
 ## Done When

package/plugins/fh-meta/agents/challenger.md

@@ -1,6 +1,8 @@
 ---
 name: challenger
 description: Frontier-grade adversarial evaluator for harness assets, papers, designs, and code. Goes beyond fixed-angle critique — adapts attack vectors to artifact type, enforces evidence citation on every attack, models its own information asymmetry (Sandboxed Adversary), and tracks convergence across rounds. Returns structured [issue · location · severity] output consumable by steel-quench, harvest-loop, and sim-conductor. Use when you need adversarial pressure that a self-reviewing author cannot generate.
+tools: Read, Grep, Glob, WebSearch, WebFetch
+version: 0.1
 ---
 
 > **Dual registration**: ships in `plugins/fh-meta/agents/challenger.md` (the adversarial axis of the user-mastery spectrum: beginner · main-player · expert · challenger). External plugin installs get it directly — no hub clone required.
@@ -151,7 +153,11 @@ Cross-session mode eliminates accumulated session context — use when the main
 
 ## Integration Hooks
 
-**steel-quench Wave 1** *(planned)*: challenger is designed to replace or supplement the devil agent. When wired, S-grade output feeds into Wave 2 defense round. Currently steel-quench calls `fh-commons:quench-challenger` for Wave 1 — explicit challenger wiring is a future integration step.
+**steel-quench Wave 1** *(partially wired — 2026-08-11)*: challenger is **already listed in the
+deployed persona panel** (`steel-quench/SKILL_detail.md` §T0 Claude row: `challenger · quench-challenger
+· expert`), so "planned" was stale. What is still absent is an explicit
+`Agent(subagent_type="challenger")` dispatch call — Wave 1 invokes `fh-commons:quench-challenger`.
+Panel membership ≠ a call site; do not read the first as the second.
 
 **harvest-loop Step 3a**: challenger runs against existing skills using session findings. S-grade attacks on existing skills → HIGH synthesizer grade. LOW-confidence attacks on new proposals → MED grade (defer pending verification).
 

package/plugins/fh-meta/agents/expert.md

@@ -1,6 +1,6 @@
 ---
 name: expert
-description: Frontier-grade domain-authority evaluator. Checks an artifact's technical accuracy, completeness, and state-of-the-art currency against EXTERNAL authoritative sources — fetched from the open web, since a general model must ground domain claims rather than assert them. Top tier of the user-mastery spectrum (beginner → main-player → expert): the professor / prolific author / frontier-harness operator. Every accuracy judgment carries an external citation; ungrounded assertions are withheld. Returns parallax-compatible output. Use when "is this technically correct and current with the field?" matters.
+description: 'Frontier-grade domain-authority evaluator. Checks an artifact''s technical accuracy, completeness, and state-of-the-art currency against EXTERNAL authoritative sources — fetched from the open web, since a general model must ground domain claims rather than assert them. Top tier of the user-mastery spectrum (beginner → main-player → expert): the professor / prolific author / frontier-harness operator. Every accuracy judgment carries an external citation; ungrounded assertions are withheld. Returns parallax-compatible output. Use when "is this technically correct and current with the field?" matters.'
 tools: Read, WebSearch, WebFetch
 version: 0.1
 ---

package/plugins/fh-meta/agents/fact-checker.md

@@ -18,7 +18,13 @@ The main agent passes you:
 2. The relevant scope (which directories, which memory files, which CATALOG sections)
 3. Any specific suspicions (e.g., "I think §X already covers this" or "memory may be stale")
 
-If scope is not specified, default scope = entire hub (`knowledge/`, `tracks/`, `memory/`, `CATALOG.md`).
+If scope is not specified, default scope = `knowledge/`, `tracks/`, `CATALOG.md`.
+**Check each scope element exists before searching it, and report a missing one — never fold it into
+a CLEAR.** Measured 2026-08-11: the previous default also listed `memory/`, which does not exist at
+the repo root (the memory index lives outside the repo, unreachable from `Read/Grep/Glob` without an
+absolute path), and `tracks/` is gitignored so a fresh clone has ~8 files in it. A scope element that
+resolves to nothing produces the exact failure this agent's own self-check names — *"Missing scope =
+false CLEAR"* — so emit `scope-missing: <path>` in the verdict block and mark the verdict PARTIAL.
 
 ## Two definitions of "fact-check"
 

package/plugins/fh-meta/agents/hub-persona-auditor.md

@@ -51,7 +51,8 @@ Each proposal: 1-2 lines, concrete (which section, what change).
 
 ### 5. Verdict
 
-One sentence: `SHIP` / `SHIP_AFTER_🟥` / `REVISE` — and the single most important reason.
+One sentence: `SHIP` / `SHIP_AFTER_MANDATORY` / `REVISE` / `OUT_OF_SCOPE` — and the single most important reason.
+(Verdict tokens are ASCII and greppable on purpose: an emoji inside a token — the old `SHIP_AFTER_🟥` — breaks any consumer that matches on it, and `OUT_OF_SCOPE` is emitted below but was missing from this enum, so a reader could not tell it was a legal verdict rather than an error string.)
 
 ## Operating rules
 

package/plugins/fh-meta/agents/main-player.md

@@ -91,7 +91,10 @@ Cross-tier conflicts: [where serving one tier hurts another — e.g., defaults g
 ## Integration Hooks
 
 - **sim-conductor Area A / D-code** — main-player is the engaged-use persona (A-2); Heavy tier supplies the edge-case lens for code artifacts.
-- **install-doctor adjacency** — Heavy H5 (conflicts/duplication/overwrite) complements install-doctor; main-player reports the *user-experienced* symptom, install-doctor the structural cause.
+- **install-doctor adjacency** *(one-way — install-doctor does not know about this agent; measured
+  2026-08-11: `grep -rc main-player` in that skill → 0)* — Heavy H5 (conflicts/duplication/overwrite)
+  complements install-doctor; main-player reports the *user-experienced* symptom, install-doctor the
+  structural cause.
 - **challenger boundary** — challenger *attacks* edges adversarially; main-player reports edges as a real heavy user *experiences and depends on* them. Different vantage, intentionally.
 
 ## Done When

package/plugins/fh-meta/agents/persona-innovator.md

@@ -69,7 +69,13 @@ Additional checklist for the human operator:
 
 ### 1-a. Load naming history
 
-**Path A (hub environment with naming history)**: Read `MEMORY.md` (hub memory index), then load the naming-relevant files referenced there.
+**Path A (hub environment with naming history)**: read the naming history from `CATALOG.md` and
+`knowledge/shared/` (both repo-root, both present). *Corrected 2026-08-11: this said `MEMORY.md`,
+which does not exist at the repo root (`ls MEMORY.md` → No such file; control: `CATALOG.md` and
+`README.md` resolve). Path A therefore read nothing and rendered as "no naming gap found" — and
+Path B could not catch it, because its condition is "external environment" and this IS the hub.*
+**If neither source resolves, degrade to Path B and say so in Section 0 — a silent empty read is
+not a finding of zero.**
 
 **Path B (external environment)**: Skip memory read. Use only the naming pattern taxonomy below (§ Naming pattern taxonomy) and the current invocation context.
 
@@ -77,7 +83,9 @@ Additional checklist for the human operator:
 
 Scan current asset inventory using Grep/Glob:
 ```
-grep -r "candidate\|gap\|unnamed\|no name" <harness-root>/
+Grep tool: pattern `candidate|gap|unnamed|no name`, path `.` (repo root)
+    — stated as a Grep-tool call, not a shell line: this agent declares `tools: Read, Grep, Glob`
+      and has no Bash, so a shell command here is an instruction it cannot execute.
 ```
 
 Also look for:

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

@@ -2,7 +2,7 @@
 name: agent-composer
 description: Reads the current work context and plans the optimal agent dispatch. Clarifies direction with 1-2 questions when unclear; infers and proceeds immediately when execution path is unclear. Runs an automatic recording gate after each Wave completes. Triggered by "compose agents", "which agent should I use?", "run in parallel", or "agent-composer".
 user-invocable: true
-allowed-tools: ["Read", "Bash", "Glob", "Grep"]
+allowed-tools: ["Read", "Bash", "Glob", "Grep", "Agent", "Skill"]
 model-note: session-inherit — Sonnet base is first-class (sonnet_floor_doctrine.md); depth-critical judged steps route to dispatch (opus agent / cross-family sidecar, consent-gated), never a substrate requirement
 ---
 
@@ -150,7 +150,7 @@ This orchestrator does not read files or understand structure directly — **eve
 
 Default composition table by task type.
 
-> **Note**: This table lists known installed agents. Capability fit scoring in Step 0.2 overrides static mapping when `agent_cards.json` has more current data.
+> **Note**: This table lists known installed agents. Capability fit scoring in Step 0.2 overrides static mapping when `.claude/registry/agent_cards.json` (repo-relative — there is no bare `agent_cards.json` at the repo root) has more current data.
 
 > **Call method distinction**: `(S)` = Skill tool call / `(A)` = Background dispatch via Agent tool
 

package/plugins/fh-meta/skills/apex-review/SKILL.md

@@ -177,6 +177,11 @@ Incorporate sim-conductor results into the HTML deck and save the revised versio
 All steps 0–4 completed
 + Verdict (✅/⚠️/❌) output for all target decision-makers
 + Gate verdict (Passed / Conditionally passed / Rejected) stated
++ Deck artifact exists on disk: `apex_review_deck_YYYYMMDD.html` (Step 2) — confirm by printing its
+  path and byte size; this is the skill's only durable output, so "reviewed but no deck" is NOT done
++ If the gate verdict is `Conditionally passed`: the `/sim-conductor Area E` verdict has been
+  RECEIVED and folded in (the mandatory-next gate below). Without that verdict the run is not done,
+  regardless of the A/B/C/D selection
 + User next step selection (A/B/C/D) confirmed
 ```
 

package/plugins/fh-meta/skills/asset-placement-gate/SKILL.md

@@ -2,7 +2,7 @@
 name: asset-placement-gate
 description: Routes a proposed skill, plugin, or agent to its correct home — forge-harness (FH) meta-skill, project-local agent, or drop — by applying a 4-criteria meta-skill bar followed by a project-local value test.
 user-invocable: true
-allowed-tools: ["Read", "Grep", "Glob"]
+allowed-tools: ["Read", "Grep", "Glob", "Bash"]
 model: sonnet
 ---
 
@@ -73,19 +73,37 @@ Criterion ④ ("no overlap with existing FH skills") is otherwise an LLM **recal
 ground truth — a duplicate skill with a novel name passes because the judge has no enumerated list to
 check against (judge-robustness swarm, 2026-06-13). Ground it mechanically first:
 
+Run this in **Bash** from the FH repo root (the pipeline below is `grep | grep -v | grep -c`, which the
+Grep tool cannot express).
+
 ```bash
-# enumerate existing skill names + descriptions (grounds the judged comparison)
-grep -riE "name:|description:" plugins/fh-meta/skills/*/SKILL.md plugins/fh-commons/skills/*/SKILL.md
-# hard-collision check: WHOLE proposed name or a WHOLE trigger phrase reused verbatim.
+# LEG A — enumerate existing skill names + descriptions (grounds the judged comparison).
+# This leg is ALSO the liveness check for leg B: a dead scan and a genuine no-collision both
+# produce collision-count 0, so the enumerate count must be read FIRST.
+ROSTER="$(grep -riE 'name:|description:' plugins/fh-meta/skills/*/SKILL.md plugins/fh-commons/skills/*/SKILL.md)"
+ROSTER_N="$(printf '%s' "$ROSTER" | grep -c . )"
+echo "roster_entries=$ROSTER_N"
+if [ "$ROSTER_N" -eq 0 ]; then
+  echo "SCAN_DEAD — enumerate leg returned 0 rows (wrong cwd / missing plugins tree / glob did not match)."
+  echo "Criterion ④ = UNDETERMINED. Do NOT read this as 'no collision'. Fix the cwd and re-run."
+  # fail-closed: stop here, do not run leg B, do not pass ④
+fi
+# LEG B — hard-collision check (run ONLY when ROSTER_N > 0).
+# WHOLE proposed name or a WHOLE trigger phrase reused verbatim.
 # grep -wF (whole-word, fixed-string) on the full strings — NOT -E on tokens (a shared common
 # word like "review" is not a collision). Exclude the asset's own file (self-match = false hit).
-SELF="plugins/fh-meta/skills/<proposed name>/SKILL.md"
-grep -rwF -e "<proposed full name>" -e "<full trigger phrase 1>" -e "<full trigger phrase 2>" \
+SELF="plugins/fh-meta/skills/PROPOSED_NAME/SKILL.md"
+grep -rwF -e "PROPOSED_FULL_NAME" -e "FULL_TRIGGER_PHRASE_1" -e "FULL_TRIGGER_PHRASE_2" \
      plugins/*/skills/*/SKILL.md | grep -v "$SELF" | grep -c .
 ```
 
-Surface **collision count + nearest existing skill(s)**. Criterion ④ then passes only if **0 whole-name/
-whole-trigger collision AND the judged ≤90%-overlap check agrees**. A verbatim whole-name or
+(Replace the `PROPOSED_*` / `FULL_TRIGGER_*` tokens with the literal strings under evaluation.)
+
+Surface **roster_entries + collision count + nearest existing skill(s)** — the roster count is part of
+the report, not a private step: `collision=0` is only meaningful next to a non-zero `roster_entries`
+(not-found ≠ zero). Criterion ④ then passes only if **`roster_entries` > 0 AND 0 whole-name/
+whole-trigger collision AND the judged ≤90%-overlap check agrees**. `roster_entries = 0` is
+`SCAN_DEAD` → ④ **UNDETERMINED** (fail-closed), never a pass. A verbatim whole-name or
 whole-trigger reuse is a hard ④ fail regardless of the LLM judgment. **Honest scope**: the grep grounds
 *literal* name/trigger reuse only — a post-cutoff duplicate with a *paraphrased* trigger is invisible to
 both the judge (cutoff) and the grep (literal); that residual leans on the judged half **fed the
@@ -105,10 +123,22 @@ skipping. (Provenance: `tracks/_audit/session_2026_07_25_claude5-context-rules-s
 
 ## Done When
 
+Each condition declares its check class (mandatory-pass / measured / judged); every judged condition
+names its adversarial pairing — no judge-only path.
+
 ```
 All steps 0–3 completed
+  (mandatory-pass — each step's output block is present; a skipped step is a FAIL, not a default pass)
++ Step 0.5 scan is LIVE: roster_entries > 0
+  (measured: the reported roster_entries count. 0 = SCAN_DEAD → criterion ④ UNDETERMINED and this
+   skill is NOT done — a dead scan and a genuine no-collision both read 0 collisions, so the
+   liveness number is what separates them)
 + Step 3 routing result output (location: FH meta-skill / local agent / drop)
+  (judged — adversarial pairing: `fh-meta:challenger` re-argues the case for the destination that was
+   NOT chosen, citing the 4 criteria; a routing verdict that survives the opposite case passes, an
+   unopposed one does not)
 + Next action specified (write SKILL.md / create .claude/agents/ / none)
+  (mandatory-pass — a literal next-action string from that enum; blank or "TBD" does not satisfy it)
 ```
 
 ---

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

@@ -103,11 +103,25 @@ probe localhost:11434 && echo "ollama-local(mac)"
 EP="$FH_SIDECAR_EXTRA"
 [ -z "$EP" ] && [ -f tracks/_meta/sidecar_endpoints.env ] && \
   EP="$(grep '^OLLAMA_EXTRA=' tracks/_meta/sidecar_endpoints.env | cut -d= -f2- | tr -d '"')"
-for e in $EP; do
-  case "$e" in *[!0-9a-zA-Z.:-]*|'') continue;; esac   # host:port form only — reject glob/junk (no word-split hole)
+# Split on NEWLINES via `while read`, never on an unquoted `for e in $EP`. Word-splitting an unquoted
+# variable is a BASH behavior; zsh does NOT word-split, so under zsh a space-separated multi-endpoint
+# list arrives as ONE string, hits the charclass below (space is not in it) and is dropped — silently
+# reporting "no extra sidecars", i.e. the single-family fail-open this skill exists to prevent.
+# Measured 2026-08-11: `EP="host1:11434 host2:11434"` → bash probes 2, zsh probes 0.
+# Accepts either separator: commas/spaces are normalized to newlines first.
+printf '%s\n' "$EP" | tr ' ,' '\n\n' | while IFS= read -r e; do
+  [ -z "$e" ] && continue
+  case "$e" in *[!0-9a-zA-Z.:-]*)
+    # NEVER a silent `continue` — a dropped endpoint must be visible, or an unprobed panel is
+    # indistinguishable from an unavailable one (DEGRADED_PANEL_UNUSED vs DEGRADED_SINGLE_FAMILY).
+    echo "sidecar-endpoint DROPPED (not host:port form): [$e]" >&2; continue;;
+  esac
   probe "$e" 10 && echo "ollama-extra($e)"             # -m10: a sleeping GPU box may wake slower than 6s
 done
 ```
+**Shell note**: this block is `sh`-portable and is written so bash and zsh behave identically. Any
+future edit that reintroduces `for e in $EP` re-opens the zsh drop above. A drop line on stderr is a
+**finding, not noise** — carry it into the Step 6 degrade verdict rather than discarding stderr.
 Endpoint resolution is a **mechanical env/file read** (not a prose instruction the runner must remember),
 so this discovery is tier-independent — no target-tier sim owed. The extra-endpoint binding lives only in
 the gitignored `tracks/_meta/sidecar_endpoints.env` (auto-synced to the companion store); the public skill

package/plugins/fh-meta/skills/context-doctor/SKILL_detail.md

@@ -29,13 +29,37 @@ ls -d */ 2>/dev/null | head -20
 ### Step 2 — Large file detection + warning format
 
 ```bash
-# Detect files exceeding 500 lines (top 10)
-find . -name "*.py" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" -o -name "*.java" -o -name "*.kt" \
-  | grep -v node_modules | grep -v .git \
-  | xargs wc -l 2>/dev/null | sort -rn | head -11 | tail -10
+# Detect files exceeding 500 lines (top 10).
+#
+# Two mechanical guards, both measured 2026-08-11 on this repo:
+#  (a) 500-line THRESHOLD via awk. Without it, `head -11 | tail -10` emits 10 rows unconditionally —
+#      this repo's largest source file is 450 lines and ZERO files exceed 500, yet the old pipeline
+#      still returned 10 rows, which the warning block below then rendered as "Large file detected".
+#      Not-found must render as "none", never as a hit.
+#  (b) `while read` instead of `xargs wc -l`. xargs splits a long file list into batches and EACH
+#      batch emits its own `total` row; those rows then occupy slots in the reported top 10
+#      (measured: 6000 files -> 2 `total` rows, one of them inside the top 10). The loop also
+#      survives spaces in filenames, which the xargs form did not.
+FILES="$(find . -type f \( -name "*.py" -o -name "*.ts" -o -name "*.tsx" -o -name "*.js" \
+                        -o -name "*.java" -o -name "*.kt" \) \
+              -not -path "*/node_modules/*" -not -path "*/.git/*")"
+SCANNED="$(printf '%s' "$FILES" | grep -c . )"
+BIG="$(printf '%s\n' "$FILES" | while IFS= read -r f; do
+         [ -n "$f" ] || continue
+         printf '%s\t%s\n' "$(wc -l < "$f")" "$f"
+       done | awk -F'\t' '$1 > 500' | sort -rn | head -10)"
+
+if [ "$SCANNED" -eq 0 ]; then
+  echo "large-file scan: UNMEASURED — 0 source files matched (wrong cwd, or this project uses other extensions)"
+elif [ -z "$BIG" ]; then
+  echo "large-file scan: none — 0 of $SCANNED source files exceed 500 lines"
+else
+  printf '%s\n' "$BIG"
+fi
 ```
 
-When files exceeding 500 lines are found:
+Emit the warning below **only for rows the scan actually returned** (`BIG` non-empty). A `none` or
+`UNMEASURED` line is reported as-is and produces **no** warning — and `UNMEASURED` is not `none`:
 
 ```
 ⚠️  Large file detected: {filename} ({N} lines)
@@ -67,11 +91,22 @@ If found, suggest adding to the `## Check Items` or `## Token Efficiency` sectio
 # CLAUDE.md line count
 wc -l CLAUDE.md .claude/CLAUDE.md 2>/dev/null | sort -rn | head -3
 
-# MEMORY.md line count (200-line limit)
-wc -l memory/MEMORY.md 2>/dev/null
-
-# memory/*.md files exceeding 30K
-find memory -name "*.md" -size +30k 2>/dev/null | xargs wc -l | sort -rn | head -10
+# Session memory lives OUTSIDE the repo — under the Claude Code project dir, keyed by a slug of the
+# absolute cwd. There is NO `memory/` at the repo root: the old `wc -l memory/MEMORY.md 2>/dev/null`
+# and `find memory ... 2>/dev/null` both printed nothing on this repo, and that silence read as
+# "no bloat" (not-found rendered as zero). Derive the path; never hardcode a home path.
+MEMDIR="$HOME/.claude/projects/$(pwd | sed 's|/|-|g')/memory"
+if [ -d "$MEMDIR" ]; then
+  # MEMORY.md line count (200-line limit)
+  if [ -f "$MEMDIR/MEMORY.md" ]; then wc -l "$MEMDIR/MEMORY.md"
+  else echo "MEMORY.md: UNMEASURED — index absent under $MEMDIR"; fi
+  # topic files exceeding 30K (while-read: survives spaces, no xargs `total` rows)
+  find "$MEMDIR" -name "*.md" -size +30k | while IFS= read -r m; do
+    printf '%s\t%s\n' "$(wc -l < "$m")" "$m"
+  done | sort -rn | head -10
+else
+  echo "memory audit: UNMEASURED — no memory dir at $MEMDIR. This is NOT 'zero bloat'."
+fi
 
 # SKILL.md files > 300 lines with no SKILL_detail.md (salience-splitter candidates)
 find plugins -name "SKILL.md" 2>/dev/null | while read f; do

package/plugins/fh-meta/skills/corpus-grounding-expander/SKILL.md

@@ -14,8 +14,13 @@ that corpus from multiple public-domain versions so grounding is robust and non-
 verified in ANY version counts), without ever adding a generator.
 
 > Origin: harvested from the-bible (2026-06-20) — 6 public-domain Bible versions, 197k verses, as the
-> fail-closed grounding axiom. Generalizes to any verbatim-relay corpus (legal statute, RFC text,
-> standards). See `tracks/_contrib/field_harvest_2026-06-20_gate-locality-and-grounding-capabilities.md`.
+> fail-closed grounding axiom. **The grounds, inline, because the harvest record does not ship**: one
+> edition's wording makes a fail-closed check reject a faithful quote that followed a *different*
+> edition, so the union across versions is what stops fail-closed grounding from failing closed
+> *against the user* — and it keeps one translation's editorial choices from silently becoming the
+> axiom. Generalizes to any verbatim-relay corpus (legal statute, RFC text, standards).
+> Full harvest record — **hub-local, not distributed in the npm package**:
+> `tracks/_contrib/field_harvest_2026-06-20_gate-locality-and-grounding-capabilities.md`.
 
 ## Triggers
 - "get more sources" / "broaden the grounded corpus"
@@ -39,12 +44,16 @@ verified in ANY version counts), without ever adding a generator.
 4. **Wire grounding as a union** — the grounding check passes if the quote matches the canonical text
    at that key in ANY version. Never add a path that generates text — grounding is quote-only.
 5. **Relay-integrity check** — confirm the consumer (the gate) QUOTES the corpus and cannot emit
-   un-grounded text; run a fabrication probe (a known non-source quote must fail-closed).
+   un-grounded text; run a fabrication probe (a known non-source quote must fail-closed) and write
+   the probe's **actual output** into the index under `probe_status:`. Writing the probe *spec*
+   without running it is `probe_status: not_run`, not a pass.
 
 ## Done When
 - **Each added source carries a verifiable public-domain/license record** in the index. *Check class: mandatory-pass (binary — license field present and non-empty per version).*
-- **Every version is normalized to the same key schema** (cross-version keys align). *Check class: mandatory-pass (a shared sample key resolves in each version or is explicitly canon-scoped).*
-- **The index declares the quote-only union contract** (grounding matches verbatim text in ANY version; no generation path) and ships a fabrication-probe spec for the consumer to run. *Check class: mandatory-pass (binary — quote-only contract + probe-spec present in the index).*
+- **The index enumerates every version that was ATTEMPTED, each carrying a status** — `fetched` | `failed(<reason>)` | `skipped(<reason>)` | `unknown`. A version that was tried and dropped appears with its reason; **absence from the list is not a zero**, it is an unrecorded drop. *Check class: mandatory-pass (binary — the attempted-version list is present and every entry carries one of the four statuses).*
+- **Every version is normalized to the same key schema** (cross-version keys align). *Check class: mandatory-pass (a shared sample key resolves in each version). A key genuinely absent from a version is allowed ONLY when the index records it under `canon_scope:` — the reason plus the explicit list of affected keys and versions. An unrecorded exception is FAIL: "it's canon-scoped" asserted in prose and nowhere in the index does not satisfy this condition.*
+- **The index declares the quote-only union contract** (grounding matches verbatim text in ANY version; no generation path). *Check class: mandatory-pass (binary — quote-only contract present in the index).*
+- **The fabrication probe was RUN and its output shows fail-closed** — a known non-source quote was submitted to the consumer gate and refused, with the refusal output recorded in the index. *Check class: mandatory-pass (binary — `probe_status: executed` plus the attached refusal output). A probe **spec** with no run is `probe_status: not_run` and is a sub-condition, never this one.* **Degrade — labeled, never a silent pass**: no consumer gate exists yet → record `probe_status: no_consumer (<what is missing>)`; the corpus may still ship, but ships **labeled relay-unverified**, this condition stays UNMET, and nothing downstream may cite the corpus as relay-verified until a run is attached.
 
 ## Guards
 - **Grounding, never a generator** — the relay constraint is structural; this skill wires *grounding*.