16k-baseline-trigger-guard: refuse an unmeasured baseline comparison on every trigger

[HackTuah]

Gate - baseline may only decrease is a required status check that had been passing having compared nothing.

github.event.before is absent from the schedule and workflow_dispatch event payloads, so ci.yml expanded to ./tools/gate.sh baseline "", and baseline_gate matched the empty string in the same branch as the all-zeros sentinel: it printed no previous ref (new branch); nothing to compare — about a branch that is main — and returned 0. Every nightly run in this repository's history was green on zero comparisons, on commits whose parents carried the baseline file it declined to read.

Measured before any fix, both quoted from their job logs in FINDINGS F1/F2:

run	event	head	job	result
34028977996	schedule	778accb8	101474937084	success, 0 comparisons
34071415451	workflow_dispatch	86681a2	101589359652	success, 0 comparisons

The guard for that payload already existed in Gate - attestation, and in run 34028977996 it measured while the baseline job did not. This reuses that elif rather than deriving a second one.

What changed

• .github/workflows/ci.yml — the attestation job's guard applied to the baseline job, routing an empty-or-sentinel before to ${{ github.sha }}~1. The on: comment claiming the schedule is "NOT for the baseline gate" is corrected; this diff falsified it in both halves.
• tools/gate.sh — baseline_gate fails closed on an empty ref. The all-zeros sentinel keeps its pass as its own labelled case. The header comment claimed "Fails closed on zero comparisons" while the empty-ref path did not.
• apps/hacktui_core/test/ci_baseline_guard_test.exs (new) — pins both halves. An ordinary hard-blocking test, not a new Gate - job, so no ruleset change (CLAUDE.md §4c). It asserts the baseline job's entire body against a literal, splits lines the way YAML does, and runs the gate rather than reading it.
• tools/mutants/c5.tsv — the rows that prove the pin bites. Several are reviewers' own survivors; HANDOFF.md §10 names which.
• BACKLOG.md — §13 records SCR-69's still-open half; §14 the documentation-drift class; §15 what the pin covers and what it does not.
• HANDOFF.md — §10 for this slice. §1's stale figures are marked in place and deliberately not corrected.

Review

Ten rounds, twenty independent lanes, each on a checkout of the index.

The gate has been byte-identical since round 4 and no lane in any round could make it report a false green. What kept failing was the pin and the prose — one mistake at eight levels, each an assertion about something the assertion does not read. Reviewers defeated the pin with a parked early return, a second guard on a different payload key, a computed sentinel, an interpreter prefix, an extra space, a glob containing no gate.sh token, and a carriage return that makes a dropped comment execute. Each is now a mutant row that dies.

The record was cut on the owner's decision of 2026-09-07: no transcribed count survives in a tracked file — every number is either the command that derives it or it is absent — because five consecutive rounds had blocked on a stale one, twice by the same mechanism. The final two lanes verified the closing delta was additive-free by exact preimage reconstruction.

Still open, deliberately

SCR-69's other half: a force-push sets before to a well-formed but unreachable SHA, which is not the sentinel, so the gate goes red on a tree that is fine. Opposite failure direction, different remedy, recorded in BACKLOG.md §13.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VNdY6wVS773UJgrHUeBN3Y