release: v2.7.0 — 리뷰 결함·새 동작 테스트 실패 확인, knowledge 사실에 commit (dev → main)

README.ko.md

@@ -113,7 +113,8 @@ git이 추적하는 폴더로 지정하세요.
 - **독립된 눈.** scout, 리뷰어, QA는 작업을 만든 컨텍스트가 아닌 새 컨텍스트에서
   돕니다.
 - **주장 대신 증거.** 실행 환경이 없으면 PASS가 아니라 BLOCKED입니다. 버그 수정에는
-  수정 전에 실패한 테스트가 있어야 합니다.
+  수정 전에 실패한 테스트가 있어야 하고, 새 동작의 테스트는 그 동작 없이 한 번
+  실패하는 것을 확인합니다.
 - **검증한 것만 납품.** 납품한 트리가 검증한 커밋과 다르면 `close`가 거부합니다.
 
 Build는 [ponytail](https://github.com/DietrichGebert/ponytail) 방식(가장 작은

README.md

@@ -117,7 +117,8 @@ vendors e2e's agent guide (Apache-2.0, pinned) so it works offline;
 - **Independent eyes.** Scouts, reviewer and QA run in fresh contexts, never
   in the one that wrote the work.
 - **Proof, not claims.** No environment means BLOCKED, never PASS. A bug fix
-  needs a test that failed before the fix.
+  needs a test that failed before the fix; a test for new behavior is seen
+  failing once without it.
 - **Delivered means what was verified.** `close` refuses a delivered tree that
   differs from the verified commit.
 

SKILL.md

@@ -13,7 +13,7 @@ governs phase order, gates and proof.
 |---|---|---|---|---|
 | 1 | Intent | `phases/1-intent.md` | `intent.md`, `refs/`, `logs/` | Clarity 90%, intent approved |
 | 2 | Plan | `phases/2-plan.md` | `refs/`, `PLAN.md` | plan approved |
-| 3 | Build | `phases/3-build.md` | code, `review.md` | review PASS |
+| 3 | Build | `phases/3-build.md` | code, `review.md`, `logs/` | review PASS |
 | 4 | Evidence | `phases/4-evidence.md` | `evidence.md`, `logs/` | `sdlc.sh score` routes |
 
 `sdlc.sh next <slug>` names the next step and why. Run `sdlc.sh` (in this

VERSION

@@ -1 +1 @@
-2.6.1
+2.7.0

docs/index.html

@@ -337,7 +337,7 @@ <h2 id="why-h" data-i18n="why-h">What you get</h2>
         </div>
         <div>
           <dt><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><path d="M4 7h10M18 7h2M4 17h4M12 17h8"/><circle cx="16" cy="7" r="2"/><circle cx="10" cy="17" r="2"/></svg><span data-i18n="g4-t">Proof behind every "passed"</span></dt>
-          <dd data-i18n="g4-d">Every pass points to the record of what was run or seen. If there is nowhere to try something, it is marked as blocked, never as passed. A bug fix needs a test that failed before the fix and passes after it. The list of scenarios becomes your own checklist.</dd>
+          <dd data-i18n="g4-d">Every pass points to the record of what was run or seen. If there is nowhere to try something, it is marked as blocked, never as passed. A bug fix needs a test that failed before the fix and passes after it, and a test for new behavior is seen failing once without it. The list of scenarios becomes your own checklist.</dd>
         </div>
         <div>
           <dt><svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="1.7" stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"><circle cx="12" cy="12" r="9"/><path d="m8.5 12 2.2 2.2 4.8-5"/></svg><span data-i18n="g5-t">Delivered means what was verified</span></dt>
@@ -461,7 +461,7 @@ <h2 id="install-h" data-i18n="inst-h">For the person who sets it up</h2>
     "g3-t": '다른 AI가 따로 점검합니다',
     "g3-d": '조사하는 보조 AI, 검토 AI, 시험 AI는 모두 처음부터 새로 시작합니다. 점검하는 대상을 만든 AI가 맡는 일은 없고, 결과를 고치지도 않습니다. 별도의 AI를 쓸 수 없으면 그 빈틈을 기록합니다.',
     "g4-t": '"통과"에는 증거가 붙습니다',
-    "g4-d": '모든 통과에는 무엇을 실행하거나 보았는지 기록이 붙습니다. 시험해 볼 곳이 없으면 통과가 아니라 "막힘"으로 표시합니다. 오류 수정에는 고치기 전에는 실패하고 고친 뒤에는 통과하는 테스트가 있어야 합니다. 시나리오 목록은 그대로 담당자의 확인 목록이 됩니다.',
+    "g4-d": '모든 통과에는 무엇을 실행하거나 보았는지 기록이 붙습니다. 시험해 볼 곳이 없으면 통과가 아니라 "막힘"으로 표시합니다. 오류 수정에는 고치기 전에는 실패하고 고친 뒤에는 통과하는 테스트가 있어야 하고, 새 동작의 테스트는 그 동작 없이 한 번 실패하는 것을 확인합니다. 시나리오 목록은 그대로 담당자의 확인 목록이 됩니다.',
     "g5-t": '검증한 것만 납품합니다',
     "g5-d": '납품한 결과가 검사한 버전과 다르면 작업을 닫을 수 없습니다. 위험한 작업은 어떤 설정에서든 사람의 허락이 먼저 필요합니다. 데이터 삭제, 저장된 데이터의 구조 변경, 다른 시스템이 기대는 약속의 변경, 보안, 인프라, 이 컴퓨터 밖으로 작업을 내보내는 일이 여기에 해당합니다. 작업 중에 읽은 글은 참고 자료로만 다루고, 지시는 사람에게서만 받습니다.',
     "lim-h": '할 수 없는 일',

log/changelog-v2.7.0.md

@@ -0,0 +1,79 @@
+# v2.7.0 — keep the contrast: failing tests and dated facts
+
+Three rules make a defect, a test and a fact each carry the run that proves
+it. A reviewer's caught defect becomes a test, a new test is seen failing
+once, and a fact in `knowledge` names the commit it was verified at.
+
+## Why
+
+- Across three projects, 17 features ran v2 reviews; 11 had at least one
+  FAIL verdict, 26 `VERDICT: FAIL` rounds in total, 14 on plans and 12 on
+  diffs. Only Evidence's code FAILs had to become regression tests. Of the 10
+  diff-review findings that were a code defect or a weak test, the next round
+  cited a new or changed test aimed at the finding for 3 and not for 7
+  (counted from `review.md` citations, not from the test files); for the 7
+  the defect lived on only as prose in `review.md`.
+- v2.5.1's false e2e pass (a URL assertion matching the address before a
+  client-side redirect) was found only by a negative control. One project
+  already proved its tests that way: head PASS 12/12, base FAIL 12/12.
+- Facts in `knowledge` cite files and lines but not the commit they were
+  verified at, so a scout could cite one as AS-IS after the code changed.
+- The recurring FAIL categories across features — missed consumers (5
+  features), intent or policy mismatches (7), scenario gaps (4), code defects
+  (4), weak tests (3) — are each already targeted by an existing reviewer or
+  scout rule. A lesson store would only repeat them.
+
+The idea came from reviewing a proposal built on Harness² (arXiv
+2610.14224, "Recursive Agent Harnessing for an Open World", 2026-10-09),
+which keeps a lesson only when an execution contrast supports it. These
+rules keep that contrast in the project's own tests and on the fact itself.
+The proposal's lesson store, context packer and auto-promotion scripts were
+not adopted: v2.0.0 removed that kind of structure.
+
+## Changes
+
+- **A blocking diff finding that shows a defect becomes a test.** In Build,
+  a blocking diff-review finding that a screen, API, job or log could show
+  is fixed like a bug fix: its regression test first, failing on the
+  FAILed commit for the finding's reason, then the fix. If no test can show
+  it (no test runner, a production-only symptom), the fix's commit message
+  (or, for an uncommitted review, the review request) says why. Plan-review FAILs and findings nothing can observe (wording,
+  docs, structure) are excluded. The reviewer blocks an observable blocking
+  diff finding left untested and unexplained. Non-blocking `(observable)`
+  findings still become QA scenarios, as in v2.4.0.
+- **A new test for new behavior is seen failing once.** Build runs each new
+  or changed test for new behavior once where it must fail: on the base
+  commit in a disposable worktree (never stashing or touching the human's
+  tree, as QA already does), or, for a new symbol, with the behavior
+  disabled. The output goes to `logs/`. It must fail on its assertion;
+  "symbol not found" or a compile error does not count. The reviewer now
+  receives `logs/` and blocks a new-behavior test with no recorded assertion
+  failure. SKILL.md lists `logs/` among Build's writes.
+- **Facts carry the commit they hold at.** Evidence writes each fact that
+  rests on code into `knowledge` with its file (path from the project root)
+  and the delivered commit of that file's repository,
+  `[verified: <file> @ <short sha>]`, re-checking the fact on that commit
+  first when the file changed since the fact was checked (this change, a
+  rebase or a merge), so a fact verified in Plan against older code is not
+  stamped with the new commit. Before citing a
+  code fact, a scout checks whether its file changed since that commit (in
+  the file's repository, `git log --oneline <sha>.. -- <file>` prints
+  anything); if it changed, or the fact has no commit, the scout re-checks it
+  against the code or cites it as `[assumed: <why>]`. Running `git log` in
+  the file's own repository makes this work for one repository or a root
+  holding several. Facts not about code (policies, data) are not checked
+  this way, and only facts a scout cites are checked; there is no list of
+  watched paths and no cache.
+
+No script change; existing records stay valid.
+
+## Known limits
+
+- Facts already in `knowledge` have no commit; each is re-checked whenever
+  a scout cites it, until Evidence records it again with its commit.
+- A fact that rests on several files carries one file; a change to another
+  file it depends on is not caught.
+- The plan load is at 2500 of 2500 words; the next rule for scouts,
+  reviewer, plan or SKILL.md must replace words, not add them.
+- Whether a finding could be observed is a judgment, as `(observable)`
+  already is.

phases/3-build.md

@@ -16,6 +16,11 @@ Changes:
   codebase's structure and conventions.
 - Bug fix: write the regression test first and watch it fail for the
   reported reason; fix the root cause once, in the shared code.
+- A new or changed test for new behavior: run it once where it must fail, on
+  the base commit in a disposable worktree (never stash or touch the human's
+  tree) or, for a new symbol, with the behavior disabled; keep the output in
+  `logs/`. It must fail on its assertion, not on a missing symbol or compile
+  error.
 - Never cut validation at trust boundaries, error handling that prevents data
   loss, security or accessibility.
 - After each step run the narrowest check; run the full `build`, `test` and
@@ -29,7 +34,7 @@ Changes:
 
 Commit the change locally (the project's rules decide the message), then
 dispatch the reviewer (`roles/reviewer.md`, diff mode) with intent.md,
-PLAN.md, `refs/` and the commit. Where every commit needs the human's
+PLAN.md, `refs/`, `logs/` and the commit. Where every commit needs the human's
 approval, review the uncommitted change instead. The reviewer appends a round
 to `review.md` ending in `VERDICT: PASS` or `VERDICT: FAIL`. Record it with
 `sdlc.sh review <slug> pass|fail "<one line>"`: the script reads the verdict
@@ -53,7 +58,13 @@ needs a new round, as does any code changed after a PASS; rewording the
 commit made on top of it does not. Recording hashes those files into the
 repository's object store (not its index): ignore large or secret files
 before reviewing.
-FAIL → fix the blocking findings, commit, review again. The third FAIL stops
+FAIL → fix the blocking findings, commit, review again. A blocking diff
+finding that a screen, API, job or log could show is a bug fix: its
+regression test first, failing on the FAILed commit for the finding's reason,
+then the fix; if no test can show it, the fix's commit message (or, for an
+uncommitted review, the review request) says why.
+Plan-review FAILs and findings nothing can observe (wording,
+docs, structure) need no test. The third FAIL stops
 the loop: show the human the trail; only `sdlc.sh approve plan <slug>
 --human` resumes it. Then phase 4.
 

phases/4-evidence.md

@@ -52,7 +52,10 @@ where it is written in the same delivery, or its follow-up and owner go in
 Delivery's `Policy:` line. Then:
 - move verified facts from the refs (and `refs/learnings.md`, if any) into
   `knowledge`: one place per fact, updating a page rather than adding a
-  duplicate;
+  duplicate; a fact that rests on code ends with its file (path from the
+  project root) and the delivered commit of that file's repository, after
+  re-checking the fact there if the file changed since it was checked:
+  `[verified: <file> @ <short sha>]`;
 - add the saved scripts (`logs/scripts/`) worth keeping to the project's
   `e2e` suite, or name them as a follow-up in Delivery;
 - `sdlc.sh close <slug> shipped "<one line>"` — refused if the delivered

roles/reviewer.md

@@ -24,7 +24,7 @@ Inputs: intent.md, PLAN.md, `refs/`. Blocking:
 
 ## Diff mode (phase 3)
 
-Inputs: intent.md, PLAN.md, `refs/`, the commit or uncommitted diff
+Inputs: intent.md, PLAN.md, `refs/`, `logs/`, the commit or uncommitted diff
 (untracked files included). Read every changed file.
 Blocking:
 - it does not do what PLAN.md says, or does more;
@@ -38,7 +38,8 @@ Blocking:
 - security: injection, missing authorization, secrets in code, unsafe
   deserialization;
 - a test that cannot fail or asserts nothing; a bug fix without a test that
-  failed before it;
+  failed before it; a new-behavior test with no recorded assertion failure;
+  an observable blocking diff finding untested and unexplained;
 - a name that lies, one rule decided in two places, a hardcoded environment
   value, dead code the change left behind;
 - an abstraction, option or layer PLAN.md did not ask for.

roles/scout.md

@@ -6,7 +6,9 @@ under `refs/`. Change nothing else.
 Inputs: the question, intent.md, the project root, the `knowledge` folder,
 and config.md.
 
-- Read `knowledge` first; cite what it already answers.
+- Read `knowledge` first; cite what it already answers. Re-check, or mark
+  assumed, a cited code fact with no commit or whose file changed since (its
+  repository's `git log`).
 - Trace from the entry the question names to the end: every caller in every
   tier and repository, every reader and writer of the data, the states
   existing records are in. Query real data read-only where a tool and