C3D-2204: fix the review gates head-blindness

[matt-manuel]

Rolls out the head-drift fix for the review gate (C3D-2204). Workflow file only — no application code.

The defect

The gate passed on label presence alone and never read the marker's head=. It fires on synchronize — the one event that invalidates a marker — so it re-ran and confidently passed green exactly when it was most likely wrong.

The fix

When a review is recorded, compare the marker's head=<40-char sha> against the PR's current head:

state	result
marker head == PR head	pass
marker head != PR head	fail — HEAD DRIFT, naming both SHAs and the delta to review
marker carries no head=	warn and pass — "cannot verify", not "drifted"
review-exempt	pass, not head-sensitive (an exemption is about the kind of change)

Semantics are deliberately identical to ~/reviews/tools/review-gate.sh, including last head= wins so a supplemental pass re-anchors the marker. Two answers to "is this reviewed?" that disagree would be its own defect.

Break-test evidence — observed, not asserted

Extracted the run: block and executed it against a mocked gh, 10 scenarios:

label + marker head MATCHES current    rc=0 want=0  PASS
label + marker head STALE (drift)      rc=1 want=1  PASS
label + marker has NO head=            rc=0 want=0  PASS
supplemental: LAST marker matches      rc=0 want=0  PASS
supplemental: LAST marker stale        rc=1 want=1  PASS
no label, marker matches (warn+pass)   rc=0 want=0  PASS
no label, no marker                    rc=1 want=1  PASS
review-exempt ignores drift            rc=0 want=0  PASS
dependabot auto-C                      rc=0 want=0  PASS
chore(deps title auto-C                rc=0 want=0  PASS

Then sabotaged the comparison itself (if [ "$MARKER_HEAD" = "$PR_HEAD_SHA" ] → if true): the drift case flips to rc=0, confirming that comparison is what produces the red rather than something incidental.

The break test earned its keep — it caught a real defect in the fix. grep -o colourised the match, so the captured SHA carried ANSI escapes and never string-equalled the PR head: the gate reported HEAD DRIFT on every correctly-reviewed PR. Fails closed and loud, but wrong, and it only reproduces where grep chooses to colourise. Fixed with --color=never, which the sibling shell script already had for exactly this reason.

Review status — stated plainly

This has NOT had a multi-leg adversarial review. It carries the break-test evidence above and a single (orchestrator) opinion. Deliberately left unlabeled so the gate flags it rather than self-exempting — a workflow whose job is to enforce review should not wave itself through.

[matt-manuel]

Adversarial review — Tier A, 6 opinions (full write-up on c3d-mcp-server#201)

The shipped workflow is byte-identical across all four rollout PRs (sha256 13fc4c17…, verified on every head), so one cycle covers all four. The reconciled findings, false-positives section and recommendation live on https://github.com/CognitiveVR/c3d-mcp-server/pull/201 rather than being duplicated here.

Verdict: merge is fine; promotion to a required check is not. This is a strict improvement on the head-blind gate it replaces and introduces no regression. But the gate can be turned green by anyone able to post a comment (24 characters, invisible in the rendered UI, no write access needed on the two PUBLIC repos) — pre-existing, and the thing that must close before enforcement mode.

All five legs answered the planted question YES independently.