Two independent cold reviews (Codex/GPT-5.6 cross-family, and an isolated Opus reviewer).
Both returned BLOCKED. Every finding below was reproduced before being acted on.
BLOCKER — the new test file had no `docs/guard-inventory.md` row, so
`test_the_inventory_covers_exactly_the_guards_that_exist` failed and took all 22
`test_mutation_harness.py` parametrisations down with it at fixture setup: the whole
`script-tests` CI context was red. Reproduced locally (1 failed, 22 errors). The row is
added, and the hand-maintained summary count with it (21 -> 22 proof files) — that
summary is itself asserted against the table, which is the second half of the same trap.
BLOCKER (Codex), VERIFIED AGAINST THE v1.27.1 SOURCE, not fixed here — the timeline
walk's `null` terminator does not prove exhaustion.
`ListIssueCommentsAndTimeline` applies LIMIT/OFFSET in `FindComments` and filters
AFTERWARDS, dropping `CommentTypeCode` rows and inaccessible xrefs into a NIL slice that
serializes as bare `null`. So a page of inline review comments — which a PR author can
create — reads as the end of the list while later pages still hold events, and rows are
ASCENDING so the events a fence wants are furthest from page 1. This predates #803 and
defeats the BASE fence identically, and closing it is a redesign of a walk both axes
share. Filed as #870 with the source evidence; the fence is now documented as closing the
ABA on a timeline with no such truncating block, NOT the ABA outright. Adding a fourth
overclaiming contract to the change that removed three is the one outcome not acceptable
here.
Two fail-opens in that shared walk, both pre-existing, both fixed:
* an empty ARRAY first page was trusted on ANY page while the `null` arm required
`page > 1` — so `[]` on page 1 certified "nothing retargeted, nothing pushed" from a
response the walk cannot explain. The comment three lines above already asserted the
safe behaviour, so the code contradicted its own stated rule for one of the two empty
shapes this server produces.
* no row was validated before `.type` was selected on, so a page of rows with no readable
`.type` certified a zero count. `pr-changed-files.sh` has applied
validate-what-you-consume to every row it extracts since #643.
FALSE MEASUREMENT — "PR #761 carries five, three of them `true`" is wrong; it is FOUR
true and one false (ids 20485/21341/21450/21452 true, 21451 false). Re-measured and
corrected in both copies.
The self-trigger stall is REFUTED FROM THE SOURCE rather than from my 20-sample margin:
v1.27.1 creates the push comment BEFORE emitting the synchronize notification, so a run
always sees its own causative event. The measurement is kept for the margin and corrected
— `started_at` is a LOWER bound, since the job runs a checkout and several steps before
counting; timed end to end on run 2385 (PR #868) the real margin is 69s, 16.7s of it
in-job. A second, DIFFERENT inversion route is now named: a PR's first push is recorded
~1s after creation and fires no `synchronize`, so an `opened` run landing in that window
would abstain with no successor. Unreachable at 26s+ queue latency; recorded, not fixed.
Three tests were passing for the wrong reason, all found by review, all now mutation-proved:
* the unreadable-head hook test asserted only that SOME `ask` happened, and this fixture
asks anyway for an unrelated reason — so a mutation reading an unreadable head as
unchanged left all 107 merge-consent tests green.
* every synthetic `pull_push` row carried `is_force_push: true`, making "every push is
counted, the boolean is not read" untestable: narrowing the tally to force-pushes passed
the whole suite. The flag now alternates.
* every fence test terminated on the default `null`, so deleting the head count from the
empty-ARRAY arm left the head fence silently dead for `[]` timelines. Now parameterised.
And my own `.type` test passed for the wrong reason on first writing — the mode served
untyped rows on every page, running the walk into its page cap, which withholds anyway.
The overclaim sweep was incomplete: grepping the retracted WORDING missed four
paraphrases ("at one head" twice, "bound to ONE head" in the other caller of the same
script, "bound to BOTH the given sha" in a record's machine-readable mechanics line, which
left frontmatter contradicting its own body). The record no longer states a count.
New coverage: cross-page accumulation (the double could not serve a multi-page timeline at
all, so a page-1-only walk passed everything), both terminator shapes, the head arm's
trust guard, and row validation. Mutations verified to redden: delete the head arm (4),
value-not-movement (1), fold the tallies (1), page-1-only walk (3), restore the `[]`
fail-open (1), drop row validation (1), count force-pushes only (2), drop the head count
from the `[]` arm (1), drop the head arm's trust guard (1), and Codex's named hook
mutation (1).
Verification: 1117 passed, 2 skipped; ruff clean; decisions-validate OK; YAML + `bash -n`
clean. No `.cs` touched, so the BOM/format gate does not apply.
Refs #870
Decisions-Edit: yes