16d-review-binding: gate the three rules that failed under attention

[HackTuah]

Replaces #8, which was closed to clear a blocked rollup. The commit is unchanged — 27de8a1, tree 3797e134, trailer Reviewed-diff: e72d17e8…, signoff written by tools/signoff.sh over the tree both round-5 reviewers read.

What lands

Three rules produced during slice 16, each of which had already failed once while enforced by attention alone, become mechanical.

1. .githooks/pre-commit refuses a commit on main. Positive match on the full ref, so a tag named main cannot defeat it (--short returns heads/main in that case). Detached HEAD is deliberately exempt — rebase, bisect and the reviewer's index checkout all run detached. A slip-guard, not an authorization control.

2. .githooks/pre-commit greps tracked .md for every path the staged diff deletes or renames, failing on a hit not marked <!-- historical-path --> on the hit line or the one immediately above. Two lines above does not count.

   Earlier drafts failed open three separate ways, each caught by review and each fixed with a working negative control: C-quoted paths (non-ASCII, tab, quote, newline) that git grep -F could not match; blobs GNU grep calls binary, where it exits 0 and emits nothing on stdout; and an enumeration exiting 128, where the affirmative printed over an input never searched. Now NUL-delimited throughout, with both enumerations checking git's exit status.

   Scope, stated exactly: a tracked path, deleted or renamed by this commit, cited by exact string in a tracked .md. The three defects that motivated the rule are not in that domain — internal/ has never been tracked, one instance is a JSON key and one a stale line number — and the hook says so.

3. tools/signoff.sh binds the signoff to the tree the reviewers read. Exactly two reviewer tree files per round, distinct reviewer numbers, both equal to git write-tree; .githooks/pre-commit re-checks at commit time. This exists because in slice 16c three of six staged files changed on disk mid-review: the attestation covers index bytes while the drift test hashes worktree bytes.

Also: one definition of the reviewable-diff recipe in tools/gate.sh (copies 3 → 2), with a test holding .githooks/commit-msg's remaining copy token-identical until slice 17 removes it; the two tools/gate.sh comment corrections carried from 16b; and BACKLOG.md §11 recording the advisory-print is_uint duplicate.

Review

Four rounds plus a fifth of scope "delta", two independent reviewers each, read-only on checkouts of the index.

Round	Tree	Gate integrity	Record / regression	Blocking
1	ebf90b65	FAIL	PASS-with-residuals	2
2	11d2c441	FAIL	FAIL	4
3	a3b8b2ce	FAIL	PASS-with-residuals	2
4	f159ea53	FAIL	PASS-with-residuals	1
5 (scope: delta)	3797e134	CONFIRMED	CONFIRMED	0

Seven blocking findings. Seven universal quantifiers. Zero mechanism defects after round 1. Every blocking finding was a claim in a tracked file — every, cannot, all three, would have — and not one was a defect in the code those words described. The slice closed by stripping the last such claim rather than qualifying it.

Why #8 was closed rather than merged

Its rollup carried two check-runs per context — a passing pull_request run and a failing force-push run — so two required contexts each showed a failure.

Cause, proved across three runs on this branch:

Run	Head	before	Result
33700784040	d025d35	all-zeros (branch created)	success
33709449011	27de8a1	d025d350… (orphaned by force-push)	failure
33877201424	27de8a1	all-zeros (branch re-created)	success

Same job, same code, same head SHA between the last two — only before differs. tools/gate.sh:173-176 guards the all-zeros sentinel but not a well-formed SHA that is unreachable; :178 then fails closed, correctly. The CI blind spot is filed against slice 18 with this derivation.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VNdY6wVS773UJgrHUeBN3Y