@ai-dev-methodologies/rlp-desk 0.10.1 → 0.11.0

package/docs/blueprints/sv-architecture-rethink.md

@@ -0,0 +1,84 @@
+# Blueprint: Self-Verification Architecture Rethink (§7¾ escalation)
+
+**Status**: backlog (escalation candidate, not yet planned for implementation)
+**Filed by**: RC-1 patch (fix/rc1-tmux-sv-skip-and-rc2-prd-cross-us-lint, 2026-04-25)
+**Escalation trigger**: governance §7¾ Architecture Escalation
+
+## Why this is here
+
+RC-1 (tmux SV report 5-min hang) was fixed by **disabling** `generate_sv_report`
+in tmux runners and recording the disable on three channels (session-config,
+metadata.json, debug log) so traceability is preserved. That patch is honest
+but it does not answer the deeper question:
+
+> Should the Self-Verification report be produced by spawning `claude --print`
+> at all? And if not, what should replace it?
+
+The current shell implementation (`src/scripts/lib_ralph_desk.zsh:603-668`)
+spawns `claude --print` against a tmux pane with no usable TTY/stdin and waits
+for a 300-second watchdog. Even in Agent mode the cost surface is large:
+synchronous LLM call, no caching, no determinism, and no obvious story for
+running SV against historical campaigns offline.
+
+Per governance §7¾ this counts as an **architecture question**, not a patch
+target. Open it as a blueprint, gather options, and decide before touching the
+shell function again.
+
+## Options to evaluate
+
+1. **Static SV (no LLM call)**.
+   - Replace the LLM-generated report with a deterministic Node module that
+     reads `iter-*.json`, `cost-log.jsonl`, `verify-verdict.json`, etc., and
+     emits the 10-section report from templates + counts.
+   - Pros: no spawn, sub-second runtime, runs offline against archived
+     campaigns, cacheable.
+   - Cons: loses qualitative analysis ("Worker over-engineered" / "Verifier
+     rubber-stamped"). The 10 sections are not all reducible to counts.
+2. **Hybrid: static skeleton + on-demand LLM section.**
+   - Static skeleton always runs at campaign end. The qualitative sections
+     ("Worker Process Quality", "Verifier Judgment Quality", "Patterns:
+     Strengths & Weaknesses", "Recommendations for Next Cycle", "Blind Spots")
+     become an opt-in `rlp-desk sv enrich <slug>` command run from a real TTY
+     (Agent mode or interactive shell).
+   - Pros: best of both. tmux runner gets a useful baseline report; the user
+     can opt into the LLM pass when convenient.
+   - Cons: two code paths to maintain.
+3. **Move SV out of the runner entirely.**
+   - rlp-desk produces only the artifacts (already does). SV becomes a
+     dedicated `rlp-desk sv <slug>` command that the user invokes from any
+     working session.
+   - Pros: cleanest separation. Runner stays small; SV evolves independently.
+   - Cons: discoverability — users may forget to run it. Requires a small UX
+     for "you should run sv against this slug" hints in the campaign report.
+4. **Keep the LLM spawn but fix it for tmux.**
+   - Negotiate with claude CLI for a non-interactive, no-TTY mode that streams
+     stdout and exits cleanly. Possibly via `--output-format json` plus
+     explicit stdin closure.
+   - Pros: smallest surface change.
+   - Cons: depends on Claude CLI semantics that may shift between releases.
+     RC-1 already showed how brittle this is.
+
+## Open questions
+
+- What sections of the current 10-section SV are actually consumed by humans?
+  (Metric sections vs qualitative sections vs blind spots — gather telemetry.)
+- Is there a way to reuse the verifier model itself (already running per
+  iteration) to emit a "campaign-level" verdict at the end, rather than a
+  separate spawn?
+- Should SV be cross-engine consensus too, or is single-model SV enough?
+
+## Decision pending
+
+Open. The current RC-1 patch is the **interim** answer: SV is disabled in
+tmux, traceability is preserved through `with_self_verification_requested` +
+`sv_skipped_reason`, and the user gets an honest "requested but skipped"
+banner in the Campaign Report. This blueprint will be promoted to a plan when
+someone takes ownership of the architecture question above.
+
+## References
+
+- `src/scripts/lib_ralph_desk.zsh:603-668` — current `generate_sv_report` impl.
+- `src/scripts/run_ralph_desk.zsh:719-727, 758-768, 2118-2125, 2151` — RC-1 patch.
+- `src/node/reporting/campaign-reporting.mjs:412-591` — Node-side SV summary
+  builder (already pure file I/O — promising starting point for option 1/2).
+- governance §7¾ Architecture Escalation — escalation policy this entry uses.

package/docs/multi-mission-orchestration.md

@@ -0,0 +1,154 @@
+# Multi-Mission Orchestration Patterns
+
+rlp-desk runs **one mission per `run_ralph_desk.zsh` invocation**. The runner is
+intentionally single-purpose: it loads a slug's PRD, executes the per-US loop,
+writes a sentinel (`COMPLETE` or `BLOCKED`), and exits. Anything that needs to
+coordinate **multiple missions** in sequence — for example a flywheel that
+runs `axis-A → axis-B → measurement → improve` — is the responsibility of a
+**wrapper script** owned by the consumer.
+
+This document explains the contract between rlp-desk and a wrapper, with a
+small worked example.
+
+## Why no built-in chain
+
+Hard-coding mission sequences inside rlp-desk would couple the runner to a
+particular project's idea of "what comes next." Different consumers have
+different goals (fixed-length improvement campaigns, indefinite uptime, mission
+graphs branching on metrics). A wrapper layer keeps rlp-desk focused on the
+single-mission contract while letting each consumer encode its own policy.
+
+## Per-mission outputs the wrapper can read
+
+After every campaign rlp-desk writes the following artifacts under
+`<ROOT>/.claude/ralph-desk/`:
+
+| Path | Purpose |
+|---|---|
+| `memos/<slug>-blocked.md` | Sentinel for a BLOCKED outcome. First line is `BLOCKED: <us_id>`; second line is `Reason: <verdict reason>` (governance §1f BLOCKED Surfacing). |
+| `memos/<slug>-complete.md` | Sentinel for a COMPLETE outcome. |
+| `memos/<slug>-iter-signal.json` | Last worker signal (status, us_id, summary). |
+| `memos/<slug>-memory.md` | Campaign memory accumulated across iterations. |
+| `memos/<slug>-flywheel-signal.json` | When flywheel ran, the direction it picked. May contain a `next_mission_candidate` field that the wrapper can use to decide what to launch next. |
+| `logs/<slug>/metadata.json` | One-line summary of the campaign config, including `with_self_verification`, `with_self_verification_requested`, and `sv_skipped_reason` (RC-1). |
+
+**Exit-code contract (per entry point — read carefully):**
+
+| Entry point | Exit 0 | Exit 2 | Exit 1 |
+|---|---|---|---|
+| `src/node/run.mjs` (Node, agent mode) | clean COMPLETE | BLOCKED (verifier or model-upgrade-exhausted), reason on stderr | unhandled error (e.g. unknown flag) |
+| `src/scripts/init_ralph_desk.zsh` | scaffold OK | PRD cross-US lint reject (per-us mode), violations on stderr | scaffold incomplete or input error |
+| `src/scripts/run_ralph_desk.zsh` (zsh, tmux mode) | clean exit (sentinel decides COMPLETE/BLOCKED) | not used | any failure path; wrappers must inspect the sentinel files to tell COMPLETE from BLOCKED |
+
+The Node entry surfaces the BLOCKED reason on **stderr** with **exit code 2**.
+The zsh runner does **not** distinguish exit codes for COMPLETE vs BLOCKED;
+inspect the sentinel files (`memos/<slug>-{complete,blocked}.md`) instead.
+PRD lint rejection (init exit 2) is the only path that uses exit 2 in the
+zsh side.
+
+## Minimal wrapper recipe (zsh)
+
+The recipe below polls a fixed mission list, launches each one, and stops on
+the first BLOCKED. It demonstrates the contract without prescribing a policy.
+
+```zsh
+#!/usr/bin/env zsh
+set -u
+set -o pipefail
+
+ROOT="${ROOT:-$PWD}"
+DESK="$ROOT/.claude/ralph-desk"
+MISSIONS=(
+  "axis-1-baseline"
+  "axis-2-improve"
+  "axis-3-measurement"
+)
+
+for SLUG in "${MISSIONS[@]}"; do
+  # Skip missions that already finished (idempotent re-runs).
+  if [[ -f "$DESK/memos/$SLUG-complete.md" ]]; then
+    print "Skipping $SLUG — already COMPLETE"
+    continue
+  fi
+  if [[ -f "$DESK/memos/$SLUG-blocked.md" ]]; then
+    print "Stopping chain — $SLUG is BLOCKED:"
+    cat "$DESK/memos/$SLUG-blocked.md"
+    exit 2
+  fi
+
+  print "Launching $SLUG"
+  ROOT="$ROOT" \
+  WORKER_MODEL="${WORKER_MODEL:-gpt-5.5:medium}" \
+  VERIFIER_MODEL="${VERIFIER_MODEL:-opus}" \
+  VERIFY_MODE="${VERIFY_MODE:-per-us}" \
+  CB_THRESHOLD="${CB_THRESHOLD:-6}" \
+    zsh ~/.claude/ralph-desk/run_ralph_desk.zsh "$SLUG"
+  rc=$?
+
+  # zsh runner: rc tells you "did the script crash", not COMPLETE vs BLOCKED.
+  # Read the sentinel for the actual terminal state.
+  if [[ -f "$DESK/memos/$SLUG-complete.md" ]]; then
+    print "$SLUG completed cleanly"
+  elif [[ -f "$DESK/memos/$SLUG-blocked.md" ]]; then
+    print "$SLUG blocked — stopping chain"
+    cat "$DESK/memos/$SLUG-blocked.md"
+    exit 2
+  else
+    print "$SLUG ended without sentinel (rc=$rc) — stopping chain"
+    exit "${rc:-1}"
+  fi
+done
+```
+
+Three design notes:
+
+- The wrapper checks the **sentinel files first** — both before and after
+  invoking the runner. This makes re-runs idempotent (finished missions are
+  skipped) and accommodates the zsh runner's lack of a distinct
+  COMPLETE-vs-BLOCKED exit code.
+- For the **Node runner** (`src/node/run.mjs`, agent mode) the recipe can be
+  simpler: switch on `$rc` directly because Node uses exit 2 specifically for
+  blocked outcomes. The exit-code table above lists which entry point uses
+  which convention.
+- `init_ralph_desk.zsh`'s exit 2 (PRD lint reject) is the wrapper's only
+  pre-launch fail-fast signal. Treat it the same as a blocked sentinel — the
+  campaign will not start until the PRD is fixed.
+
+## Flywheel-driven dynamic chain (optional)
+
+**Emit side (rlp-desk responsibility)**: when a mission runs the flywheel
+review (`--flywheel on-fail`), the flywheel agent's signal JSON
+(`memos/<slug>-flywheel-signal.json`) MAY include an optional
+`next_mission_candidate` field — `null` for "no recommendation" or a slug
+string for "consumer should chain this slug next." The Node leader
+propagates this field into `status.json` (`status.next_mission_candidate`)
+so wrappers can poll either file. The flywheel prompt template
+(`init_ralph_desk.zsh` flywheel heredoc) and `governance.md` §7 ⑥½ both
+document the field. Field is OPTIONAL and absence is treated as `null` —
+backward-compat with prior flywheel signals.
+
+**Consumer side (wrapper responsibility)**: pick the next slug from that
+field instead of a fixed list:
+
+```zsh
+NEXT_SLUG=$(jq -r '.next_mission_candidate // empty' \
+  "$DESK/memos/$SLUG-flywheel-signal.json" 2>/dev/null)
+if [[ -n "$NEXT_SLUG" ]]; then
+  # Recurse or push onto the queue. Apply your own policy:
+  # - de-dupe against an already-launched set,
+  # - cap chain length to avoid runaway loops,
+  # - require `axis-history.json` distance to avoid revisiting.
+fi
+```
+
+`next_mission_candidate` is advisory only. Wrapper authors should still apply
+guardrails (max chain length, distance-from-history checks, manual approval
+gates) before consuming it.
+
+## Non-goals (explicitly)
+
+- A built-in `rlp-desk auto-chain --slug-prefix … --max-missions N` command is
+  **not** in scope. It would re-introduce the coupling we are trying to avoid.
+  If you want one, build it as a small wrapper and share it with the community.
+- rlp-desk does not validate mission ordering or dependency graphs. The wrapper
+  owns this policy.

package/docs/plans/rlp-desk-0.11-handoff-7fixes.md

@@ -0,0 +1,352 @@
+# rlp-desk 0.11 — Handoff Final 7-fix bundle (ralplan v3)
+
+> v3 changes: NEW-1 (bash→zsh fixture invocation) + NEW-2 (early-exit grep broadened) Architect executor follow-ups 흡수.
+> v2 changes (Architect + Critic codex iteration): PR split A/B 결정, R7 schema fallback, R8 helper-side guard, R9 reason canonicalization + edge cases, R10 normalized US extractor + quarantine (not rm), R11 early-exit grep inventory + trap, self-verification mechanical assertion 패치.
+
+## Context
+
+소비자 Final Handoff (`coordination/handoffs/2026-04-25-rlp-desk-final-status-and-handoff.md`) timestamp evidence 기반 7건 결함:
+
+| ID | Severity | 결함 | Root file |
+|---|---|---|---|
+| P0-D | HIGH | A4 fallback 83% 빈발 (worker iter-signal 누락) | `run_ralph_desk.zsh:1587-1595`, `:526-546` |
+| P1-F | MEDIUM | test-spec ≥3 tests/AC IL-4 자가모순 | `init_ralph_desk.zsh` test-spec gen + ingest |
+| P1-G | MEDIUM | partial_verify signal vocabulary 부재 | `init_ralph_desk.zsh:448-454` Signal rules + verifier |
+| P1-H | MEDIUM | blocked 시 memory.md/latest.md 미갱신 | worker prompt blocked exit hygiene |
+| P2-I | MEDIUM | block ≠ failure → contract defect silent 12-iter | `run_ralph_desk.zsh:2659` consecutive_blocks 신규 |
+| P2-J | MEDIUM | final ALL verify cross-mission us_id leak | `run_ralph_desk.zsh:2198/2425-2429` US_LIST scope |
+| P2-K | LOW | cost-log 비어있음 (tmux mode) | `lib_ralph_desk.zsh:367` write_cost_log call coverage |
+
+## PR 분할 결정 (v2)
+
+Architect 권고에 따라 **PR-A(protocol) + PR-B(runtime) 2-PR 분할** 채택. 사용자가 "단일 PR" 명시한 경우에도 R7 schema collision (R3 와 silent fallback 위험) 때문에 분리 필요.
+
+- **PR-A (protocol/contract)**: R5 + R6 + R7 + governance §1f/§7f/§7g + us017/us018/us019
+- **PR-B (runtime/state)**: R8 + R9 + R10 + R11 + governance §8/§7a + us020/us021/us022/us023
+- 자가검증 mapping 시나리오는 양 PR 모두 포함 (각 PR 의 fix 만 evaluate). 최종 self-verification (7/7) 은 PR-B merge 후 별도.
+
+단, 사용자 직접 "단일 PR" 재요청 시 single PR 로 진행하되 self-verification 시나리오를 더 강화 (per-row mechanical assertion 필수).
+
+## RALPLAN-DR
+
+**Principles** (4):
+1. **Fail loud, not silent** — A4 fallback / block-as-success / cross-mission leak / cost-log silence 모두 silent failure 패턴.
+2. **Backward-compat first** — verify_partial 신규 status 의 기존 wrapper malformed 처리 명시. test-spec lint warn-then-strict 단계 진화.
+3. **Minimal blast radius** — PR split + per-fix helper 분리. 각 fix 의 회귀는 독립 us_test.
+4. **Self-verification mechanical** — 변경 사항 X가 자가검증 시나리오 Y에서 실제 트리거되었음을 grep+exit-code 로 증명.
+
+**Decision Drivers**:
+1. consumer wrapper 가 동일 패턴(83% A4 fallback, cross-mission leak, contract defect silent loop) 재발 차단.
+2. 7-mission autonomous run 후 debug.log [FLOW] events 가 의미 있는 summary 보유 + audit log auditable.
+3. cost-log 빈 파일 = "broken logging" 분류 가능, audit pipeline 신뢰성.
+
+**Viable Options 비교**:
+
+(아래 옵션 비교 v1 과 동일하나 Critic ITERATE 흡수 패치 추가)
+
+- **R7 verify_partial schema malformed 처리 (Architect issue #2)**: `verify_partial` 인데 `verified_acs` 미존재/빈 배열 → `status='blocked'`, `reason='verify_partial_malformed'` 으로 다운그레이드. Worker autonomy 위배 차단.
+- **R8 helper-side guard (Critic R8 + Architect issue #3)**: Verifier 의 mtime check 만으로 부족. `write_blocked_sentinel` 자체에 hygiene check 추가 — memory.md/latest.md mtime 이 sentinel 작성 시각보다 오래됐으면 (즉 worker 가 hygiene update 안 했으면) sentinel JSON 에 `meta.blocked_hygiene_violated=true` 자동 첨부. Worker 가 잊어도 verifier 가 즉시 인지.
+- **R9 reason canonicalization (Architect issue #3)**: `_canonical_block_reason()` helper — hygiene wrapper prefix("hygiene_violated:", "wrapped:") strip 후 비교. R8 hygiene_violated 가 R9 counter 우회 차단.
+- **R9 edge cases (Critic R9)**: 첫-iter block / mission setup block 은 `infra_failure` reason 으로 분류된 경우 counter 증가 안 함 (mission abort 부적절). 명시 exempt.
+- **R10 normalized extractor + quarantine (Architect issue #4 + Critic R10)**: `grep -qE "^## $stale_us[: ]"` 대신 `awk '/^##[[:space:]]+(US-[0-9]+)([[:space:]:-]|$)/'` 로 정규화 추출 (PRD heading variation 대응). `rm -f` 대신 `mv` to `.sisyphus/quarantine/` (silent destructive 차단).
+- **R11 trap-based final write (Architect issue #6 + Critic R11)**: init placeholder 폐기. zsh `trap 'write_cost_log "$ITERATION" || true' EXIT` 추가 + early-exit path grep inventory 회귀로 보장.
+- **Self-verification per-row functions (Architect issue #5 + Critic Self-V)**: 단일 monolithic script 대신 7 함수 (`test_r5_a4_audit_triggered`, …) + 각 함수 내 pre/post 카운터 + grep 로 변경 함수 호출 증명.
+
+---
+
+## 해결 계획 (v2 patches highlighted)
+
+### Fix R5: P0-D — A4 fallback 추적 + worker prompt 강화
+
+**대상**:
+1. `src/scripts/run_ralph_desk.zsh:1587-1595` + `:526-546` — A4 fallback 발동 시 audit log entry 작성 (`a4-fallback-audit.jsonl`, append).
+2. `src/scripts/init_ralph_desk.zsh` worker prompt — "Step N+1 (mandatory)" 추가 + auto-generated summary penalty 명시.
+3. Verifier prompt — A4 fallback summary detection 시 verdict.meta.iter_signal_quality='auto_generated'.
+4. governance §1f — A4 ratio 권고 (per-mission < 10%).
+
+**검증 (us017) — Critic R5 patch 흡수**:
+- AC1: a4-fallback-audit.jsonl entry 작성 (zsh fixture)
+- **AC1+ (Critic R5)**: pre_count=$(wc -l a4-fallback-audit.jsonl), trigger fixture, assert post_count > pre_count + ratio 계산 정확.
+- AC2: worker prompt grep "Step N+1" + "iter-signal.json with SPECIFIC summary" 존재
+- AC3: governance §1f 에 "A4 ratio < 10%" 권고 텍스트 + 측정 방법 명시
+- AC4 (신규): Verifier prompt 에 "auto_generated" detection 문장 + meta field 명시
+
+### Fix R6: P1-F — test-spec ≥3/AC enforcement (warn default + strict opt-in)
+
+**대상**:
+1. `src/scripts/init_ralph_desk.zsh` — `_lint_test_density()` helper:
+   - PRD AC count 추출 (per-US, `^- AC[0-9]+:` regex)
+   - test-spec test count 추출 (per-US, `^### Test ` 또는 `^\*\*T-` 헤더 카운트)
+   - ratio < 3 시: WARN(default) → log_warn + audit + **init exit message 마지막에 summary 표시 (Critic R6 patch)**; STRICT(`--test-density-strict`) → exit 1.
+2. `src/scripts/run_ralph_desk.zsh` + `src/node/run.mjs` — `--test-density-strict` flag stub.
+3. governance §7f — Test Density Enforcement (WARN+STRICT decision tree).
+4. Worker prompt — "≥3 tests/AC (happy + negative + boundary) 강제" 강화.
+
+**검증 (us018) — Critic R6 patch 흡수**:
+- AC1: `--test-density-strict` 플래그 파싱 (zsh + Node)
+- AC2: WARN default — ratio<3 fixture 에서 init exit=0 + audit log entry **+ stderr/stdout 마지막 라인에 "Test density warning: US-XXX has N tests for M ACs (ratio=N/M < 3)" 메시지 포함**
+- AC3: STRICT — ratio<3 fixture 에서 init exit=1 + 동일 메시지
+- AC4: governance §7f 텍스트 정합 (Decision tree, downgrade 없음)
+
+### Fix R7: P1-G — verify_partial signal vocabulary
+
+**대상 (Critic R7 + Architect issue #2 patches)**:
+1. `src/scripts/init_ralph_desk.zsh:448` Signal rules — verify_partial + 필수 필드 명시.
+2. `src/scripts/init_ralph_desk.zsh build_verifier_prompt` 함수 (or equivalent prompt heredoc) — 정확 문장 추가:
+   ```
+   If signal status=verify_partial, evaluate ONLY verified_acs. Treat deferred_acs as out-of-scope (not fail).
+   ```
+3. `src/node/runner/campaign-main-loop.mjs` 신호 파싱 — verify_partial + verified_acs 미존재/빈 배열 시:
+   ```js
+   if (signalStatus === 'verify_partial' && (!Array.isArray(signal.verified_acs) || signal.verified_acs.length === 0)) {
+     // Downgrade to blocked
+     await writeSentinel(blockedSentinel, 'blocked', usId, 'verify_partial_malformed', { reason_category: 'mission_abort', recoverable: true, suggested_action: 'retry_after_fix' });
+     continue;
+   }
+   ```
+4. `src/scripts/run_ralph_desk.zsh:1313+` — verify_partial 동등 처리 (zsh 측 fallback).
+5. governance §7g 신규 — Signal Vocabulary Extension + malformed downgrade 명시.
+
+**검증 (us019)**:
+- AC1: Signal rules grep verify_partial + verified_acs/deferred_acs/defer_reason
+- AC2: governance §7g 정합 + malformed downgrade 명시
+- AC3: Node 파서 verify_partial→verified_acs 만 verifier prompt 전달 (behavioural fixture)
+- AC4: zsh 파서 verify_partial 인지
+- **AC5 (Architect issue #2)**: malformed fixture (verify_partial + verified_acs=[]) → blocked sentinel 작성 + reason='verify_partial_malformed' + reason_category='mission_abort'
+- **AC6 (Critic R7)**: Verifier prompt 에 정확 sentence 존재 (grep)
+
+### Fix R8: P1-H — Blocked exit hygiene + helper-side guard
+
+**대상 (Critic R8 + Architect issue #3 patches)**:
+1. `src/scripts/init_ralph_desk.zsh` worker prompt — Blocked exit hygiene 섹션:
+   > "On blocked exit (status=blocked): BEFORE writing iter-signal.json, ALWAYS append to memory.md § Blocking History `{iter, us, reason, suggested_repair}` AND update latest.md § Known Issues."
+2. **`src/scripts/lib_ralph_desk.zsh:write_blocked_sentinel` (Critic R8 patch)** — sentinel write 직전 hygiene check:
+   ```zsh
+   local hygiene_violated=false
+   local mem_file="$DESK/memos/$SLUG-memory.md"
+   local lat_file="$DESK/context/$SLUG-latest.md"
+   local now_ts=$(date +%s)
+   for f in "$mem_file" "$lat_file"; do
+     if [[ -f "$f" ]]; then
+       local f_mtime=$(stat -f %m "$f" 2>/dev/null || stat -c %Y "$f" 2>/dev/null || echo 0)
+       if (( now_ts - f_mtime > 300 )); then
+         hygiene_violated=true
+         break
+       fi
+     fi
+   done
+   ```
+   JSON sidecar 에 `meta.blocked_hygiene_violated=$hygiene_violated` 자동 첨부.
+3. `src/node/runner/campaign-main-loop.mjs` `_checkBlockedHygiene()` helper — blocked write 시 동등 검사 + analytics event.
+4. governance §1f — "5th channel: memory.md/latest.md hygiene update" 추가 (4 channels → 5 channels).
+
+**검증 (us020)**:
+- AC1: Worker prompt grep "Blocked exit hygiene" + "memory.md" + "latest.md"
+- AC2: governance §1f grep "5th channel" + "memory.md/latest.md hygiene"
+- AC3: Node helper `_checkBlockedHygiene` 정의 (grep)
+- AC4: behavioural — fixture: stale memory.md (mtime > 5min ago) → blocked sentinel JSON sidecar 의 meta.blocked_hygiene_violated=true
+- **AC5 (Critic R8)**: lib_ralph_desk.zsh write_blocked_sentinel 에 hygiene_violated 자동 첨부 grep + behavioural fixture
+
+### Fix R9: P2-I — consecutive_blocks counter + canonicalization + edge cases
+
+**대상 (Critic R9 + Architect issue #3 patches)**:
+1. `src/scripts/run_ralph_desk.zsh` 변수:
+   ```zsh
+   CONSECUTIVE_BLOCKS=0
+   LAST_BLOCK_REASON=""
+   BLOCK_CB_THRESHOLD="${BLOCK_CB_THRESHOLD:-3}"
+   ```
+2. **`_canonical_block_reason()` helper (Architect issue #3)**:
+   ```zsh
+   _canonical_block_reason() {
+     local raw="$1"
+     # Strip wrapper prefixes
+     echo "$raw" | sed -E 's/^(hygiene_violated:|wrapped:)//' | head -c 80
+   }
+   ```
+3. **Edge case exemption (Critic R9)** — `infra_failure` category 또는 첫 iter block 은 counter 증가 안 함:
+   ```zsh
+   if [[ "$reason_category" == "infra_failure" ]] || (( ITERATION <= 1 )); then
+     # Exempt from consecutive_blocks
+     LAST_BLOCK_REASON=""
+     CONSECUTIVE_BLOCKS=0
+   else
+     local canonical=$(_canonical_block_reason "$reason")
+     if [[ "$canonical" == "$LAST_BLOCK_REASON" ]]; then
+       CONSECUTIVE_BLOCKS=$((CONSECUTIVE_BLOCKS + 1))
+     else
+       CONSECUTIVE_BLOCKS=1
+       LAST_BLOCK_REASON="$canonical"
+     fi
+     if (( CONSECUTIVE_BLOCKS >= BLOCK_CB_THRESHOLD )); then
+       echo '{"reason":"consecutive_blocks","count":'"$CONSECUTIVE_BLOCKS"',"last_reason":"'"$LAST_BLOCK_REASON"'"}' | atomic_write "$DESK/.sisyphus/mission-abort.json"
+       exit 1
+     fi
+   fi
+   ```
+4. `src/node/runner/campaign-main-loop.mjs` 동등 (state.consecutive_blocks + last_block_reason + canonicalReason).
+5. governance §8 — consecutive_blocks + canonicalization + exemption 명시.
+
+**검증 (us021)**:
+- AC1: BLOCK_CB_THRESHOLD 변수 정의 (default 3)
+- AC2: zsh same-reason counter logic
+- AC3: governance §8 텍스트 정합
+- AC4: behavioural — 3회 동일 reason BLOCK 후 mission-abort.json 생성
+- **AC5 (Architect issue #3)**: `_canonical_block_reason` helper 정의 + hygiene_violated prefix strip 검증
+- **AC6 (Critic R9)**: 첫-iter block exempt fixture (ITERATION=1, reason="setup_fail") → CONSECUTIVE_BLOCKS=0 유지
+- **AC7 (Critic R9)**: infra_failure category exempt fixture → CONSECUTIVE_BLOCKS=0 유지
+
+### Fix R10: P2-J — Cross-mission us_id leak + normalized extractor + quarantine
+
+**대상 (Critic R10 + Architect issue #4 patches)**:
+1. `src/scripts/init_ralph_desk.zsh` mission init — stale us_id detect + scrub:
+   ```zsh
+   if [[ -f "$SIGNAL_FILE" ]]; then
+     stale_us=$(jq -r '.us_id // empty' "$SIGNAL_FILE" 2>/dev/null)
+     if [[ -n "$stale_us" && "$stale_us" != "ALL" ]]; then
+       # Critic R10: normalized US extractor
+       prd_us_list=$(awk 'match($0, /^##[[:space:]]+(US-[0-9]+)([[:space:]:-]|$)/, m) { print m[1] }' "$PRD_FILE" 2>/dev/null | sort -u)
+       if ! echo "$prd_us_list" | grep -qx "$stale_us"; then
+         # Architect issue #4: quarantine, not rm
+         mkdir -p "$DESK/.sisyphus/quarantine"
+         mv "$SIGNAL_FILE" "$DESK/.sisyphus/quarantine/iter-signal.$(date +%s).json"
+         log "  Cross-mission stale us_id ($stale_us) — quarantined to .sisyphus/quarantine/"
+       fi
+     fi
+   fi
+   ```
+   단, BSD awk match() 3-arg 미지원 → `match() + RSTART/RLENGTH + substr()` pattern 또는 `grep -oE` + 후처리 사용:
+   ```zsh
+   prd_us_list=$(grep -oE '^##[[:space:]]+US-[0-9]+([[:space:]:-]|$)' "$PRD_FILE" 2>/dev/null | grep -oE 'US-[0-9]+' | sort -u)
+   ```
+2. `src/scripts/run_ralph_desk.zsh:2425-2429` final ALL verify scope — US_LIST 만 신뢰 (signal_us_id US_LIST 에 없으면 무시 + warn).
+3. `src/node/runner/campaign-main-loop.mjs` — 동등 처리.
+4. governance §7a — cross-mission us_id leak 방어 + quarantine path 명시.
+
+**검증 (us022)**:
+- AC1: init 단계 stale us_id detect + quarantine helper (grep + behavioural)
+- AC2: zsh runner final ALL verify US_LIST 신뢰
+- AC3: governance §7a 텍스트 정합 + quarantine path
+- AC4: behavioural — fixture mission PRD (US-001~003) + stale signal us_id=US-005 → SIGNAL_FILE quarantine 이동, .sisyphus/quarantine/ 에 파일 존재
+- **AC5 (Architect issue #4)**: rm -f 사용 안 함 (`grep -n "rm -f.*SIGNAL_FILE" src/scripts/init_ralph_desk.zsh` = 0)
+- **AC6 (Critic R10)**: PRD heading variation fixture (`## US-005 -`, `## US-005:`, `## US-005`) → 모두 정상 인식 (false positive 0)
+
+### Fix R11: P2-K — Cost log non-empty + trap-based final write + early-exit inventory
+
+**대상 (Critic R11 + Architect issue #6 patches)**:
+1. `src/scripts/lib_ralph_desk.zsh:367` write_cost_log — note 필드 (bytes=0 시 'no_actual_usage_recorded').
+2. **`src/scripts/run_ralph_desk.zsh` (Architect issue #6)** — main loop 진입 직후 trap 등록:
+   ```zsh
+   trap '_emit_final_cost_log' EXIT
+   _emit_final_cost_log() {
+     [[ -n "${ITERATION:-}" ]] && [[ "${COST_LOG_FINAL_WRITTEN:-0}" -eq 0 ]] && {
+       write_cost_log "$ITERATION" 2>/dev/null || true
+       COST_LOG_FINAL_WRITTEN=1
+     }
+   }
+   ```
+3. **Early-exit path inventory (Critic R11 + Architect NEW-2)** — us023 회귀가 다음 broadened grep 결과의 모든 path 가 trap coverage 내인지 검증:
+   ```bash
+   grep -nE '^[[:space:]]*(exit\b|return\b|die\b)' src/scripts/run_ralph_desk.zsh src/scripts/lib_ralph_desk.zsh | grep -v '^[^:]*:[^:]*:.*\${' > early_exits.txt
+   ```
+   `die` wrapper 함수가 `lib_ralph_desk.zsh` 에 정의된 경우 명시적으로 trap 우회 분석 + 회귀에 포함.
+4. (init placeholder 삭제 — Architect issue #6) — 빈 cost-log 가 "broken logging" 으로 감지되도록 normal path 만 보강.
+5. governance §7 Cost Tracking — tmux estimated path + trap 명시.
+
+**검증 (us023)**:
+- AC1: write_cost_log 에 note 필드 (bytes=0 시 'no_actual_usage_recorded')
+- AC2: zsh runner 에 `trap '_emit_final_cost_log' EXIT` 존재 (grep)
+- AC3: behavioural — write_cost_log 호출 후 cost-log.jsonl 비어있지 않음
+- **AC4 (Critic R11)**: early-exit grep inventory + 모든 path 가 trap coverage 검증 (스크립트 내 모든 `exit N` 또는 `return N` 위치 grep + trap fire 시점 비교)
+- **AC5 (Architect issue #6)**: init placeholder 코드 부재 (grep `placeholder.*cost-log` = 0)
+
+---
+
+## 자가검증 시나리오 — Mechanical per-row (v2)
+
+`tests/test_self_verification_0_11_handoff.sh` — 7 함수 + 각 함수 내 pre/post + grep 증명:
+
+```bash
+test_r5_a4_audit_triggered() {
+  local audit="$LOGS_DIR/a4-fallback-audit.jsonl"
+  local pre=$(wc -l < "$audit" 2>/dev/null || echo 0)
+  # Trigger: simulate done-claim without iter-signal
+  echo '{"us_id":"US-001","status":"complete"}' > "$DESK/memos/${SLUG}-done-claim.json"
+  rm -f "$DESK/memos/${SLUG}-iter-signal.json"
+  # NEW-1 (Architect): zsh fixture invocation (run_ralph_desk.zsh is zsh, NOT bash)
+  # us017 implementation MUST extract A4 fallback into a callable helper in lib_ralph_desk.zsh
+  # so it can be sourced cleanly. Until then, use zsh -c with explicit DESK/SLUG/ITERATION exports.
+  zsh -c "DESK='$DESK' SLUG='$SLUG' ITERATION=1 LOGS_DIR='$LOGS_DIR' source src/scripts/lib_ralph_desk.zsh; _emit_a4_fallback_audit US-001 1" 2>/dev/null
+  local post=$(wc -l < "$audit" 2>/dev/null || echo 0)
+  [[ "$post" -gt "$pre" ]] || { fail "R5 A4 audit not triggered (pre=$pre post=$post)"; return 1; }
+  # Mechanical: grep that the patched code path was exercised
+  grep -q "a4_fallback" "$audit" || { fail "R5 audit entry missing"; return 1; }
+  pass "R5 A4 fallback audit triggered ($pre→$post)"
+}
+
+test_r6_test_density_warn() {
+  # Fixture: PRD with 3 ACs, test-spec with 1 test
+  local stderr_capture=$(./init_ralph_desk.zsh --slug test-r6 --prd fixtures/r6-bad-prd.md 2>&1)
+  echo "$stderr_capture" | grep -q "Test density warning" || { fail "R6 init exit message missing warning"; return 1; }
+  pass "R6 test density warning emitted to stderr"
+}
+
+# ... R7~R11 동일 패턴: 각 함수가 (1) pre-state 캡처, (2) 변경 코드 직접 invoke, (3) post-state grep 검증
+```
+
+| Fix | 시나리오 | Mechanical 증명 |
+|---|---|---|
+| R5 P0-D | done-claim 작성 + iter-signal 누락 → A4 fallback 발동 | `wc -l a4-fallback-audit.jsonl` pre/post 비교 + entry grep |
+| R6 P1-F | test-spec AC 3개 + test 1개 fixture | stderr 의 "Test density warning" 라인 grep |
+| R7 P1-G | iter-signal status=verify_partial fixture (정상 + malformed) | verifier prompt grep `verified_acs only` + malformed → blocked sentinel meta.reason='verify_partial_malformed' |
+| R8 P1-H | blocked sentinel + memory.md unchanged 5min+ | sentinel JSON sidecar `meta.blocked_hygiene_violated=true` jq 추출 |
+| R9 P2-I | 동일 reason 3회 BLOCK + canonicalization + edge cases | mission-abort.json 존재 + jq `.count==3` + first-iter exempt fixture CONSECUTIVE_BLOCKS=0 검증 |
+| R10 P2-J | PRD US-001~003 + stale signal us_id=US-005 + heading variation | `.sisyphus/quarantine/iter-signal.*.json` 존재 + 원본 SIGNAL_FILE 부재 + 3 variation fixture 정상 인식 |
+| R11 P2-K | tmux mode 5 iter run + early-exit fixture | `cost-log.jsonl` 행 수 ≥ 5 + 모두 note 필드 보유 + trap fire 검증 |
+
+**Pass criterion**: 7/7 mechanical 증명 + 각 fix 가 변경된 함수/파일을 실제 호출했음을 grep 으로 확인 (tautology 방지).
+
+---
+
+## 변경 대상 파일 표
+
+```
+src/scripts/init_ralph_desk.zsh        # R5(worker prompt), R6(test density lint + flag), R7(Signal rules + verifier prompt), R8(blocked exit hygiene), R10(stale us_id quarantine)
+src/scripts/run_ralph_desk.zsh         # R5(A4 audit), R6(--test-density-strict), R7(verify_partial parsing), R9(consecutive_blocks + canonical + exempt), R10(US_LIST scope), R11(trap)
+src/scripts/lib_ralph_desk.zsh         # R8(write_blocked_sentinel hygiene_violated), R11(write_cost_log note + bytes=0 path)
+src/node/run.mjs                       # R6(--test-density-strict stub)
+src/node/runner/campaign-main-loop.mjs # R7(verify_partial parser + malformed downgrade), R8(_checkBlockedHygiene), R9(consecutive_blocks state), R10(stale us_id scrub)
+src/governance.md                      # R5(§1f A4 metric), R6(§7f Test Density), R7(§7g Signal Vocabulary + malformed), R8(§1f 5th channel), R9(§8 cb + canonicalization + exempt), R10(§7a quarantine)
+
+[테스트]
+tests/test_us017_a4_fallback_audit.sh
+tests/test_us018_test_density.sh
+tests/test_us019_verify_partial.sh
+tests/test_us020_blocked_hygiene.sh
+tests/test_us021_consecutive_blocks.sh
+tests/test_us022_cross_mission_us_leak.sh
+tests/test_us023_cost_log_nonempty.sh
+tests/test_self_verification_0_11_handoff.sh   # mechanical per-row
+```
+
+## 검증 (Self-Verification Gate)
+
+1. **LOW** — `zsh -n` + `node --check` (~10s)
+2. **MEDIUM** — us017~us023 7 신규 회귀 (~3min)
+3. **CRITICAL** — us001/us007/us012/us013/us014/us015/us016 회귀 무손실 (~3min)
+4. **자가검증 매핑 시나리오** — `test_self_verification_0_11_handoff.sh` 7/7 mechanical 증명
+
+## 단일 PR 진행 결정 (사용자 명시 시)
+
+사용자가 PR split 거부 + 단일 PR 명시한 경우:
+- R5+R6+R7 (protocol) + R8+R9+R10+R11 (runtime) 단일 PR
+- self-verification 시나리오는 양 영역 모두 포함하므로 보장 유지
+- 단, codex review iteration 5+ 도달 시 split fallback 자동 트리거
+
+## ADR (간결)
+
+- **Decision**: 7건 fix. v2 patches: PR split 권고 (사용자 명시 시 단일), R7 schema fallback (verify_partial_malformed downgrade), R8 helper-side hygiene check, R9 canonical reason + edge exempt, R10 normalized extractor + quarantine, R11 trap-based final write + early-exit inventory, self-verification mechanical per-row.
+- **Drivers**: silent failure 가시화 + backward-compat + minimal blast radius + mechanical self-verification.
+- **Alternatives considered (각 R 별 v1 표 + v2 새 patches)**.
+- **Consequences**: PR-A 먼저 머지 + soak → PR-B (권고). 단일 PR 도 가능. Worker prompt 길이 약간 증가. test-spec WARN 다수 발생 가능 (점진 strict 화).
+- **Follow-ups**: test-density STRICT 의 default 화 (v0.12+), verify_partial deferred_acs 자동 우선 재시도, A4 fallback 0% 시 hard fail.