# Evidence for job 3723: `research/qc/questions.js` (findings #128, #136, #138, #139, #2576, #2631) | what | sha256 | bytes | |---|---|---| | served text (base of this revision) | `1d2d785efde04af26a5531f2375bc0f2d705f6a9600634aac162660d0712206f` | 26657 | | **revised text (this return)** | `eeaf28829ff0e2d8bfdd85f63444f3a54b0367984628ac855b66983933738325` | 28082 | | `research/qc/corpus.js` (checked, unchanged -- **exactly** the #1715 revision #2631 names) | `b4051ac2d1bc44f1d1ee5720a3d37e882b9c264f8764d246ff9f9a1318b81d63` | 14012 | | `research/qc/checks.js` (checked: no `pickNewest` case in it) | -- | 97952 | Nine hunks: the file header, three comment blocks, two call sites, the generated preamble and both `registry-drift` descriptions. `questions.js` is a **library module**: it has no OUTPUT block, no `code-sha256`/`out-sha256` fingerprint (verified absent) and nothing embedded to re-embed -- its CLI prints one line to stdout and writes `research/QUESTIONS.md`. ## #136 / #2631 -- the git-date keys - l.335 (summarise): `pickNewest(g, fileLastDates(), C.rel)` -> `pickNewest(g, fileLastDates(), slash)`. - parity(): `dates.get(rel)` -> `dates.get(slash(r.file))`, with a comment naming the cause. **#2631's precondition is verified, not assumed:** the served `corpus.js` is exactly `b4051ac2d1bc44...`, the #1715 revision the finding names, and its `rel` is `path.relative(ROOT, p).split(path.sep).join('/')` -- already `/`-normalised. So this switch is a no-op on every platform and is taken as the defence in depth #2631 offers, not as the root-cause fix (which #1715 already landed). **#136's test clause could not be performed as written.** `research/qc/pick-newest-check.js` does not exist in the served corpus: `GET /docs/research/qc/pick-newest-check.js` and `/docs/research/pick-newest-check.js` both return 404, and `research/qc/checks.js` (97952 bytes) contains no `pickNewest` case. Rather than invent a file inside a revision job, I wrote the equivalent case as an ad-hoc harness, `pick_newest_harness.js`, and ran it against the **revised** module: ``` $ node pick_newest_harness.js ok win32-style keys fall back to path order: TODO.md ok slash keys pick the newest date: questions.js ok slash and C.rel agree here (no behaviour change): true ok a tie resolves to the last in path order: TODO.md ok a newer date wins over path order: questions.js ok ...and the win32 keys would have picked the other one: TODO.md ok no dates -> last in path order: TODO.md all harness cases pass ``` Case 1 is the defect #136 describes -- with backslash keys every entry is missed, both notes count as unknown/new, and the newest rule silently becomes path order; case 2 is the same fixture with the file's `slash`. Case 3 is the control: on this platform `path.relative` already emits `/`, so the switch preserves behaviour. Adding that case to a permanent check file is left to whoever owns the qc test layout -- it is a new artefact, not a fix to this one. ## #128 / #138 / #2576 -- scope stated, and the count kept out of the index `registry-drift`'s description now scopes the check to what it actually compares -- **section 2's rows** -- in both branches, and the dated six-row list in its comment is marked as a **2026-09-13 SNAPSHOT** with a new `SCOPE OF THIS CHECK` paragraph naming the parts it does *not* compare (prose, section 1's columns, the preamble, the unindexed list). The generated preamble ends its display-rule sentence with ``Display dates: ${fileLastDates().size ? 'git log' : 'no git: path order'}.`` -- the **kind only**, as #2576 requires, because `fileLastDates().size` counts every path git has ever seen under `research/` and would change on any such commit. The **count** lives in `registry-drift`'s description instead: `... (N rows; display dates: git log, M files)` / `no git: path order`. **This is why #128 was answered by narrowing, not by widening.** #128 offers two options: compare the full rendered text, or narrow the description to section 2. The full-text option would fire a `registry-stale` finding after every commit that adds a research path once the preamble carries a count -- exactly the failure #2576 describes -- so the description is scoped to the rows `rowsOf` compares, and the preamble carries no count. **#138's optional MIXED-cell suggestion is deliberately not taken**, for the same reason: changing a generated row's content (`MIXED (all PARTIAL; ...)`) would make the committed `research/QUESTIONS.md` rows stale against the generator and fire `registry-row-stale` for those ids until the index is regenerated. Both #128 and #138 are answered without touching row content. ## #139 -- author date, not commit date `fileLastDates` uses `--format=%as`, the **author** date. Every site that called it a "last-commit date" now says **last author date**: the file header, the `summarise` comment, `fileLastDates`' own comment, the generated preamble line #2576 names (formerly "newest last-commit date"), and -- beyond the three sites #139 names -- `fileAddDates`' comment ("First AUTHOR date") plus the cross-reference in `fileLastDates` ("first-author dates"). `%as` is kept; switching to `%cs` would change which note supplies a shared id's verdict, which no finding asks for. ## Verification performed - every served/revised/corpus hash re-checked at the server immediately before the upload; - `node --check questions.revised.js` -> clean; `diff -u` -> 9 hunks, no other line moves; - `grep` -> 0 remaining `last-commit date`, `Last-commit date`, `First-commit date`, `first-commit dates`; - `node research/qc/questions.js` in a minimal mirror (`research/qc/{questions,corpus}.js` + `research/QUESTIONS.md`) -> `research/QUESTIONS.md: 0 questions from 0 indexed notes, 0 unindexed`, exit 0; - the harness above (case for #136). - The corpus-wide CLI run is not reproduced here: it needs the full repository and its git log. No stdout comparison to a committed index is claimed, because this module's stdout is one line and the index is the artefact -- and its rows are unchanged by these edits. ## Unresolved obligations 1. **#136's requested test case is not added to a check file** -- the file it names does not exist. The equivalent case is in `pick_newest_harness.js` (ad hoc, this run). 2. `research/QUESTIONS.md` is not revised here (single-path revision). Its rows are unchanged, so it stays clean under the narrowed comparison; its preamble picks up the new line at the next `node research/qc.js --index`. 3. #139's alternative -- switching to `%cs` -- was not taken (above). 4. 17 of @Benjaminsen's returns wait for a verdict; nothing for the person to do.