docs(records): row 33 landed, cold-start failure reopened (#1508)

Co-Authored-By: Claude Opus 5.5 <[email protected]>

agents/darkwing/work/queue-33-review-r1-2026-10-04.md

@@ -0,0 +1,78 @@
+# Queue row 33, round 1 review (#1508)
+
+Darkwing, 2026-10-04. Request: #1508 comment 26648. Brief:
+`docs/plans/2026-10-04_queue-follow-ups.md`. Candidate:
+`agents/filbert/work/queue-33/candidate-manifest.sha256`, digest
+`fe7da3ef063c3274a4b1c780e1b1641d9511bfebfba8b8539b07566f471751e8`, base
+3e39a26a.
+
+Verdict: approve.
+
+## Checks
+
+- The manifest hashes to fe7da3ef. All 10 listed files match it in the
+  canonical working tree. Sage's request says 11 paths, but the manifest
+  has 10 lines; nothing is missing from what Filbert describes.
+- Scratch clone at 930d2757 with the 7 candidate files: `node --test
+  packages/queue/tests/` passes 148/148. `bash scripts/test-queue.sh`
+  gives 27/0; the live checks skip there, as in an export.
+- Fresh `git archive` export of 3e39a26a with the candidate files: 148/148
+  on the first, cold run, and `test-queue.sh` 27/0.
+- `scripts/mosaic queue verify --current` in the canonical checkout, with
+  the candidate's validator: ok at rev 47. The live log replays under the
+  calendar check.
+
+## The five items
+
+1. Genesis. `parseMigrationMap` has one caller, `genesis` in store.mjs,
+   and it runs after the retry check. A genesis retry still returns its
+   receipt, and `verify` never parses the map, so a pre-rule queue stays
+   readable. C1 is right.
+2. Assign wording matches the brief.
+3. HEAD moved. `head_moved` runs before each canary and again when the
+   clean run fails. The step-7 check now catches a move that update-ref
+   used to catch. Both exit without publishing, so C4's changed tests
+   assert the stronger behaviour. C2 ("another commit landed") is right:
+   the check fires for any commit. C3, the second check, closes the window
+   between the check and the hook run, and its test hits that window.
+4. Calendar. I parsed every hour 00 to 99 with minutes and seconds 0, 59,
+   60 and 99 in V8. Only 24:00 parses without round-tripping, and it moves
+   the date, so M8 is equivalent, as build.md says. For dates, V8 rolls
+   seven 2027 values (02-29, 02-30, 02-31, 04-31, 06-31, 09-31, 11-31);
+   the round trip refuses each. Year 0000 and 9999-12-31T23:59:59.999Z
+   round-trip and pass. Every time the validator reads goes through
+   `checkTime`: lines 252, 262, 281, 297, 330, 355, 356 and 1113.
+5. Cold runs: 20 recorded, all 148/148. The brief's rule closes the item.
+
+## My mutants
+
+Ten, on the candidate. Seven are killed:
+- no NaN guard (NaN reaches `toISOString` and throws a RangeError, not a
+  refusal);
+- the time message saying "date";
+- the round trip skipped for fields that allow "unknown";
+- no HEAD check before the canary;
+- no recheck after a failed clean run;
+- the step-7 check moved before commit-tree;
+- the rerun hint dropped.
+
+Three survive:
+- `Date.parse(v)` on the bare date in place of the `T00:00:00.000Z`
+  suffix. Equivalent: a bare ISO date parses as UTC midnight.
+- The genesis rule checking only `reviewers[0]`. Every new test puts the
+  owner first. The code uses `includes`, so this isn't a defect.
+- `head_moved` passing whenever `HEAD^` isn't H, so a move of two or more
+  commits slips through. Every test moves HEAD by one commit. The code is a
+  plain equality, so this isn't a defect either; I wrote the mutant to probe
+  the tests, not because the code could take that form.
+
+## Non-blocking
+
+- n1. A genesis map case with the owner second in the reviewer list
+  (`["filbert", "darkwing"]` for row 9) would kill the `reviewers[0]`
+  mutant. Optional.
+- n2. The 20 cold runs ran one at a time on an idle machine. My original
+  141/1 failure came from a fresh clone, but I don't know the machine's
+  load at that moment. If it was load-dependent, idle runs wouldn't show
+  it. The brief's rule is met, and I agree with closing the item. If the
+  failure comes back, capture the test name and the load.