leerness 1.35.11 → 1.35.12

package/CHANGELOG.md

@@ -1,5 +1,16 @@
 # Changelog
 
+## 1.35.12 — 2026-07-06 — codex 협업 라운드 ②: clean-room 평가 문서 갱신 + codex 정직성 리뷰 8건 반영 (UR-0011)
+
+**사용자 지시 "codex와 협업" 라운드 ② — clean-room 평가 문서 갱신.** `docs/clean-room-evaluations.md` 가 1.19.0(selftest 231)에서 멈춰 stale → 이번 세션 자체 적대적 헌트(verify-claim FP/gate/contract)를 추가하고, codex(gpt-5.5 xhigh)를 **정직성 감사관**으로 병렬 실행해 과장 표현을 교정.
+
+### 갱신
+- 평가를 **"에이전트 클린룸 평가"(behavior-only)** 와 **"메인테이너 자체 적대적 헌트"(self-administered, 가장 덜 독립적)** 로 명확히 구분 — 1.35.9/10/11 헌트를 정직 outcome(FP 1건 발견·수정 / 0 product bug / codex 9건 중 3확정·1환각기각)으로 표에 추가. 한계 섹션에 contract heuristic 한계(UR-0018) + scan secrets 포지셔닝 추가.
+- **codex 정직성 리뷰 8건 전부 반영**(협업 검수 결과): "every evaluation"→"most", "real development"→"development exercises", "Independently"→"Separately", "External reviews"→"maintainer-run(external models≠외부 인간)", "Codex independent review"→"separate Codex review", "converged on 0 bugs"→"0 bugs found in these probes", 그리고 핵심 — "reliably catches ... without false-failing"의 **자기모순 교정**(문서 자신이 1.35.9에서 FP를 찾았으므로 "no false failures"는 보장이 아니라 검사 중인 속성이라 명시).
+
+### 검증
+- selftest **271** (기존 clean-room 문서 가드를 정직성 마커 강화: clean-room + heuristic + **self-administered** + 재현 레시피). full e2e (문서-only 변경, 무영향). 문서 변경이라 런타임 동작 불변.
+
 ## 1.35.11 — 2026-07-06 — codex 협업 contract verify 적대적 헌트: FP 3종 수정 ($ 필드 정규식 + bracket export + 코드펜스 예제)
 
 **사용자 지시 "codex와 협업" 라운드 ① — contract verify 적대적 헌트.** codex(gpt-5.5 xhigh 독립 분석)와 Claude(프로브+맹신X 검수) 병렬로 FN/FP/crash 사냥. codex 9건 제기 → 재현으로 검증: 3건 확정 수정, 3건 by-design/heuristic 한계(백로그), 3건은 이미 수정한 `$` 정규식 계열, 1건은 codex 환각(재현 기각).

package/README.md

@@ -125,7 +125,7 @@ MIT
 <!-- leerness:project-readme:start -->
 ## Leerness Project Harness
 
-이 프로젝트는 Leerness v1.35.11 하네스를 사용합니다. AI 에이전트는 작업 전 `leerness handoff`로 컨텍스트를 적재하고, 작업 후 `leerness check`/`leerness audit`/`leerness session close`를 수행해야 합니다.
+이 프로젝트는 Leerness v1.35.12 하네스를 사용합니다. AI 에이전트는 작업 전 `leerness handoff`로 컨텍스트를 적재하고, 작업 후 `leerness check`/`leerness audit`/`leerness session close`를 수행해야 합니다.
 
 ### 정체성 — AI 에이전트 운영 레이어 (UR-0030)
 
@@ -179,7 +179,7 @@ leerness memory restore decision <date|title>
 
 ### MCP server (외부 AI 통합)
 
-Leerness v1.35.11는 stdio JSON-RPC MCP server를 내장합니다 — Claude Code · Cursor · Codex CLI 등 외부 AI에 **86개 도구**를 노출:
+Leerness v1.35.12는 stdio JSON-RPC MCP server를 내장합니다 — Claude Code · Cursor · Codex CLI 등 외부 AI에 **86개 도구**를 노출:
 
 ```jsonc
 // 카테고리별
@@ -200,7 +200,7 @@ Leerness v1.35.11는 stdio JSON-RPC MCP server를 내장합니다 — Claude Cod
 `<<autonomous-loop-dynamic>>` 신호만 보내면 AI가:
 1) 다음 라운드 후보 선정 → 2) 코드 변경 → 3) stress-v* 신규 작성 + 누적 회귀 → 4) e2e 219/219 → 5) npm pack + git tag + GitHub release → 6) main 자동 push (1.9.140+) → 7) session close → 8) 다음 라운드 예약.
 
-현재 누적: **70 라운드 (1.9.40 → 1.35.11)** · 매 라운드 GitHub release/태그 생성 · _reports/는 비공개 보존.
+현재 누적: **70 라운드 (1.9.40 → 1.35.12)** · 매 라운드 GitHub release/태그 생성 · _reports/는 비공개 보존.
 
 ### 성능 가이드 (1.9.140 측정)
 
@@ -238,6 +238,6 @@ leerness release pack --close --auto-main-push
 - `.harness/session-handoff.md`: 다음 세션 인수인계 (자동 작성)
 - `.harness/lessons.md` / `decisions.md` / `rules.md`: 영구 메모리 (5 surface)
 
-Last synced by Leerness v1.35.11: 2026-07-06
+Last synced by Leerness v1.35.12: 2026-07-06
 <!-- leerness:project-readme:end -->
 

package/bin/leerness.js

@@ -32,7 +32,7 @@ const { _evidenceQuality, _parseEvidenceStats, _shellGuardAnalyze, _claimFileInG
 // 1.9.295 (UR-0025 4단계): 정적 데이터 카탈로그 모듈 분리 (비파괴, require-based).
 const { CAPABILITY_SURFACE, POWERFUL_COMMANDS, ADAPTERS, REUSE_CATEGORIES, REUSE_CHECKLIST, _DEFAULT_PLATFORM_CONSTRAINTS, _DEFAULT_DOMAIN_CATALOG, _TOOL_CATALOG, _LSP_LANG_PATTERNS, OPTIMISM_PATTERNS, BUILT_IN_PERSONAS, STRINGS, BUILTIN_CATALOG, ROADMAP_STATUS_LABEL, ROADMAP_STATUS_COLOR, SECRET_PATTERNS, MERGE_OVERWRITE_FILES, MINIMAL_SKIP_KEYS, REQUIRED_WORKSPACE_FILES, KEYWORD_STOPWORDS, SKILL_CATALOG_PRESETS } = require('../lib/catalogs');  // 1.9.344/368/369 (UR-0025): catalog 분리 · 1.11.4 (UR-0007): _TOOL_CATALOG
 
-const VERSION = '1.35.11';
+const VERSION = '1.35.12';
 
 // 1.9.290 (UR-0037, Codex gpt-5.5 #4 수렴): CLI 전용 부작용은 require 시 실행하지 않는다.
 //   이전: warning listener 제거 / NODE_OPTIONS 변경 / chcp IIFE 가 top-level 즉시 실행 → require('harness') 시 호스트 프로세스 오염.
@@ -3823,11 +3823,12 @@ function _selfTestCases() {
         && eq(F([]), [])
         && eq(F(['a.css', 'b.md', 'c.js', 'd.test.js']).length, 2);          // 최대 2개 클러터 방지
     } },
-    { name: 'GPT-5.5 평가 #5 (1.19.1, UR-0009): 클린룸 평가 문서 공개 + 한계 명시 (행위)', run: () => {
+    { name: 'GPT-5.5 평가 #5 (1.19.1, UR-0009) + 정직성 calibration (1.35.12): 클린룸 문서 공개 + 한계 명시 + self-administered 라벨 (행위)', run: () => {
       const dp = path.join(path.dirname(__filename), '..', 'docs', 'clean-room-evaluations.md');
       if (!exists(dp)) return true;  // 패키지에 없으면 스킵(설치본 안전)
       const d = read(dp);
-      return /AI clean-room evaluations/i.test(d) && /heuristic, not semantic/i.test(d) && /npm i leerness@/.test(d);
+      // 1.35.12: 정직성 마커 강화 — clean-room 프레이밍 + heuristic 한계 + self-administered(독립 아님) + 재현 레시피.
+      return /clean-room/i.test(d) && /heuristic, not semantic/i.test(d) && /self-administered/i.test(d) && /npm i leerness@/.test(d);
     } },
     { name: 'CLI 영어화 Phase 1 (1.20.2, UR-0010): _uiLang 해석(flag>env>manifest>ko) + 첫화면 _t 적용 (행위+소스)', run: () => {
       const save = process.argv; const saveEnv = process.env.LEERNESS_LANG;

package/docs/clean-room-evaluations.md

@@ -1,67 +1,96 @@
-# Clean-room evaluations / 클린룸 평가 기록
-
-> **What this is, honestly.** These are **AI clean-room evaluations** — independent agents installing the
-> published npm package into fresh temp directories and exercising it from behavior only (no access to the
-> source tree), including adversarial attempts to defeat leerness's own verifier. They are **not**
-> third-party human security audits or peer-reviewed benchmarks. We publish them so the README claim
-> "verified by independent clean-room evaluations" is checkable rather than a marketing line.
->
-> 정직하게 말하면: 아래는 **AI 클린룸 평가**입니다 — 독립 에이전트가 게시된 npm 패키지를 빈 임시 폴더에
-> 설치해 **소스 접근 없이 행위만으로** 검증하고, leerness 검증기 자체를 무력화하려는 적대 시도까지 포함합니다.
-> 제3자 인간 보안 감사나 동료심사 벤치마크가 **아닙니다**. README 의 "독립 클린룸 평가로 검증" 주장을
-> 확인 가능하게 만들기 위해 공개합니다.
-
-## Methodology / 방법론
-
-Every evaluation follows the same shape:
-
-```bash
-W=$(mktemp -d); cd "$W"; npm init -y
-npm i leerness@<version>            # the PUBLISHED package, not the working tree
-LB="node node_modules/leerness/bin/leerness.js"
-# ... drive the CLI by behavior, assert on exit codes + output, clean up ...
-```
-
-- **Published-artifact only** — installs from npm, never reads the repo source. This is what an external user gets.
-- **Behavior-asserted** — checks process exit codes and stdout/stderr, not internal state.
-- **Adversarial** — a portion of every pass actively tries to make a *false* "done" claim pass verification
-  (comment-only stubs, empty-export shells, `assert(true)` fake tests, inflated test counts, language tricks).
-- **Both directions (맹신 X / "trust nothing")** — confirmed real defects AND rejected false alarms; a finding
-  counts only if reproduced on the published package.
-
-## Evaluations run / 수행한 평가
-
-| # | Round | Scope | Outcome |
-|---|---|---|---|
-| 1 | Universal-harness field study (5-axis) | Real Python / Node / Rust development, agent-swap handoff, adversarial attack on the verifier | Verdict "conditionally yes"; found 5 gaps (2×P1, 3×P2) — all closed in 1.17.2–1.17.6, shipped as the 1.18.0 stable minor |
-| 2 | 1.18.0 re-verification (published) | Independently re-ran the 5 gaps against the **published** 1.18.0 | 4 closed; surfaced a new P1 (Windows `--test-cmd python` blocked → false FAIL) → fixed in 1.18.1 |
-| 3 | 1.18.2 adversarial stub workflow | Bypass-hunters + false-positive-hunters vs the empty-shell stub detector | 9 one-keyword bypasses found (Object.freeze, async fn, inline comment, …) and closed; **0 false positives** across ~45 legitimate patterns |
-| 4 | 1.19.0 published-artifact re-verification | Installed `leerness@1.19.0` from npm; smoke-tested headline fixes | `--test-cmd python`, empty-shell rejection, `lens`, selftest 231 — all pass on the published package |
-
-## What these evaluations do NOT establish / 한계 (정직)
-
-leerness's verification is **heuristic, not semantic**. These evaluations show it reliably catches *obvious
-false-done claims* (missing files, empty shells, fake/unlinked tests, inflated counts) without false-failing
-honest work. They do **not** prove:
-
-- the implementation satisfies the actual requirement, or business logic is correct;
-- tests are sufficient or well-designed;
-- absence of security vulnerabilities or production-runtime correctness.
-
-Known heuristic gaps tracked for an AST/token-based redesign: multi-arg call-expression empty objects
-(`Object.assign({}, {})`), Python `def …: pass` / `...` / `raise NotImplementedError`, multi-language empty
-bodies. **Code coverage / mutation testing and third-party reproduction remain open** (the e2e is a
-spawn-based runner, which standard coverage tools do not instrument cleanly).
-
-## Reproduce it yourself / 직접 재현
-
-```bash
-W=$(mktemp -d); cd "$W"; npm init -y && npm i leerness@latest
-LB="node node_modules/leerness/bin/leerness.js"
-$LB init "$W" --yes
-# fake "done": claim an implementation that does not exist
-$LB task add "Implement X"
-$LB task update T-0001 --status done --evidence "x.js done, 5 tests passed"
-$LB verify-claim T-0001     # exit 1 — claim rejected (x.js missing, no tests ran)
-$LB selftest                # core-function integrity self-check
-```
+# Clean-room evaluations / 클린룸 평가 기록
+
+> **What this is, honestly.** These are **AI evaluations run by the project itself** — an agent installs the
+> published npm package into fresh temp directories and exercises it from behavior only (no access to the
+> source tree), including adversarial attempts to defeat leerness's own verifier. They are **self-administered**
+> (the maintainer's AI, sometimes cross-checked by a second model such as Codex), **not** third-party human
+> security audits or peer-reviewed benchmarks, and **not** independent in the "unaffiliated reviewer" sense.
+> We publish them so the README claim "checkable by clean-room evaluation" is verifiable rather than a
+> marketing line — you can re-run the recipe below yourself.
+>
+> 정직하게 말하면: 아래는 **프로젝트가 직접 수행한 AI 평가**입니다 — 에이전트가 게시된 npm 패키지를 빈 임시
+> 폴더에 설치해 **소스 접근 없이 행위만으로** 검증하고, leerness 검증기 자체를 무력화하려는 적대 시도까지
+> 포함합니다. 이는 **자기-수행**(메인테이너의 AI, 때때로 Codex 같은 2차 모델이 교차검증)이며, 제3자 인간 보안
+> 감사·동료심사 벤치마크가 **아니고**, "무관한 외부 리뷰어"라는 의미의 독립도 **아닙니다**. README 의 "클린룸
+> 평가로 확인 가능" 주장을 검증 가능하게 만들기 위해 공개하며, 아래 레시피로 직접 재현할 수 있습니다.
+
+## Methodology / 방법론
+
+Most evaluations follow this general shape:
+
+```bash
+W=$(mktemp -d); cd "$W"; npm init -y
+npm i leerness@<version>            # the PUBLISHED package, not the working tree
+LB="node node_modules/leerness/bin/leerness.js"
+# ... drive the CLI by behavior, assert on exit codes + output, clean up ...
+```
+
+- **Published-artifact only** — installs from npm, never reads the repo source. This is what an external user gets.
+- **Behavior-asserted** — checks process exit codes and stdout/stderr, not internal state.
+- **Adversarial (both directions)** — many passes include attempts to make a *false* "done" claim pass
+  verification (comment-only stubs, empty-export shells, `assert(true)` fake tests, inflated test counts) **and**
+  to make an *honest* claim wrongly fail (false-positive hunting). A finding counts only if reproduced on the
+  published package; agent-proposed findings are re-derived, not trusted (맹신 X / "trust nothing").
+
+## Agent clean-room evaluations / 에이전트 클린룸 평가 (behavior-only, published artifact)
+
+| # | Round | Scope | Outcome |
+|---|---|---|---|
+| 1 | Universal-harness field study (5-axis) | Python / Node / Rust development exercises, agent-swap handoff, adversarial attack on the verifier | Verdict "conditionally yes"; found 5 gaps (2×P1, 3×P2) — all closed in 1.17.2–1.17.6, shipped as the 1.18.0 stable minor |
+| 2 | 1.18.0 re-verification (published) | Separately re-ran the 5 gaps against the **published** 1.18.0 | 4 closed; surfaced a new P1 (Windows `--test-cmd python` blocked → false FAIL) → fixed in 1.18.1 |
+| 3 | 1.18.2 adversarial stub workflow | Bypass-hunters + false-positive-hunters vs the empty-shell stub detector | 9 one-keyword bypasses found (Object.freeze, async fn, inline comment, …) and closed; **0 false positives** across ~45 legitimate patterns |
+| 4 | 1.19.0 published-artifact re-verification | Installed `leerness@1.19.0` from npm; smoke-tested headline fixes | `--test-cmd python`, empty-shell rejection, `lens`, selftest — all pass on the published package |
+
+Maintainer-run multi-model reviews (external models — GPT-class + Claude — driven against a published build and
+reproduced before fixing) have also shaped many rounds; the confirmed findings are tracked in CHANGELOG under
+the "Nth 외부평가 / 외부리뷰" entries. "External" here means the *models*, not unaffiliated human reviewers.
+
+## Maintainer adversarial hunts / 메인테이너 자체 적대적 헌트 (self-administered, published artifact)
+
+These are the **least independent** kind — the maintainer's AI hunting its own flagship surfaces, cross-checked
+with Codex where noted. They are listed for transparency, not as third-party validation. Every finding was
+reproduced on the published package; every fix re-verified after publish.
+
+| Version | Surface | Method | Outcome (honest) |
+|---|---|---|---|
+| 1.35.9 | `verify-claim` declared-pass gate | 12-probe FP/FN matrix (single / `--json` / `--all` / `gate --claims` × reporter/growth/lenient) | 1 real false-positive found (a non-test `N/M passing` ratio wrongly gated) and fixed; 0 bypasses |
+| 1.35.10 | `gate` (the CI guardrail) | 8-probe matrix (clean/placeholder/secret/false-done/missing-file) | **0 product bugs**; 2 probe-flagged "FN"s were probe artifacts, refuted by reproduction; added a gate-level security regression guard |
+| 1.35.11 | `contract verify` (spec ↔ impl) | Claude probes + a separate Codex (gpt-5.5) review, cross-checked | Codex raised 9 candidates → **3 confirmed & fixed** (`$`-field regex FP, bracket-export FP, code-fence-example FP), **1 refuted as a hallucination**, the rest documented heuristic limits |
+
+The pattern worth noting: two of the three surfaces had **0 product bugs found in these probes** only after a
+dedicated false-positive hunt, and in each round a portion of the AI-proposed findings did **not** reproduce and
+were dropped. That is the point of publishing this — the checks are fallible, the probe matrices are finite, and
+the checks are themselves under adversarial test, including their own false-alarm behavior.
+
+## What these evaluations do NOT establish / 한계 (정직)
+
+leerness's verification is **heuristic, not semantic**. These evaluations suggest it catches many *obvious
+false-done claims* (missing files, empty shells, fake/unlinked tests, inflated counts), with no false failures
+in the tested cases — though the false-positive hunts have themselves found and fixed a real false-positive
+(1.35.9), so "no false failures" is a property under test, not a guarantee. They do **not** prove:
+
+- the implementation satisfies the actual requirement, or business logic is correct;
+- tests are sufficient or well-designed;
+- absence of security vulnerabilities or production-runtime correctness.
+
+Known heuristic gaps tracked for an AST/token-based redesign (UR-0016): multi-arg call-expression empty objects
+(`Object.assign({}, {})`), Python `def …: pass` / `...` / `raise NotImplementedError`, multi-language empty
+bodies; and for `contract verify` (UR-0018) a field/function present only in a **comment or string** still
+passes the grep-presence check, and a function exported as a **non-function value** passes (contract checks
+*presence*; `verify-claim` checks *substance* — a deliberate split). `scan secrets` is a convenience guard, not
+a replacement for a dedicated scanner (gitleaks/trufflehog). **Code coverage / mutation testing and third-party
+reproduction remain open** (the e2e is a spawn-based runner, which standard coverage tools do not instrument
+cleanly).
+
+## Reproduce it yourself / 직접 재현
+
+```bash
+W=$(mktemp -d); cd "$W"; npm init -y && npm i leerness@latest
+LB="node node_modules/leerness/bin/leerness.js"
+$LB init "$W" --yes
+# fake "done": claim an implementation that does not exist
+$LB task add "Implement X"
+$LB task update T-0001 --status done --evidence "x.js done, 5 tests passed"
+$LB verify-claim T-0001     # exit 1 — claim rejected (x.js missing, no tests ran)
+$LB selftest                # core-function integrity self-check (exit 0 when the install is intact)
+```

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "leerness",
-  "version": "1.35.11",
+  "version": "1.35.12",
   "description": "Leerness: 비파괴 마이그레이션, 자동 버전 감지·업데이트, 계획/진행/핸드오프 자동화, 게으름·시크릿·인코딩 자동 가드, Claude Code 슬래시 통합을 갖춘 한국어 우선 AI 개발 하네스.",
   "keywords": [
     "leerness",