[remigiusz-antczak]

Wave 6, assembled from independently reviewed lanes into one commit (fc265cede) on main.

What changes

1. Decision-council surface (refs veche: unvalidated --option names make seats fail silently (COULD_NOT_DECIDE) unless they are short tokens #2136, veche: a seat reply with a brace in the surrounding prose, or with scores as a second JSON object, is dropped as unparseable #2137, veche: add --json-out PATH, and say so on stderr when the gate skips the council (the twin is silently ignored) #2139, veche: make the decision record explain itself (verdict vs tally, duplicated dissent, near-duplicate kill criteria) #2140). veche.py --help states the --option name rule, the real skip gate behind --stakes and --reversibility, and how --json-out writes. --json-out writes through a temp file in the same directory and os.replace on every outcome, so a reader or a crash never sees half a record; a read-only file is replaced, while a read-only directory or a bind-mounted file exits 2. A --json-out write that fails at the end keeps exit 3 for COULD_NOT_DECIDE and exit 4 for NOT_CONVENED. A decided record carries kill_criteria_by_lens, each seat's own wording before near-duplicates collapse. The kill-criteria dedupe key folds the Unicode comparison signs, so one tripwire written both ways is kept once. references/council.md states the reply contract a custom --cmd must meet, as the parser accepts it, and scripts/test_veche_doc_claims.py pins that wording and the help text.
2. Every test wired into CI (refs Audit and retire stale guard rails and tests #1592). scripts/ci-gates.sh tests-wired fails when a scripts/test_*.py or scripts/test-*.sh file is run by no CI step and has no row with a reason in scripts/ci-test-allowlist.tsv; a stale row also fails, so the list only shrinks. Its first run found 71 tests no CI step ran: 59 now run in four suites, and the 12 corpus re-scores of 30 s or more are allowlisted with their measured runtimes. Four of the newly registered tests timed a guard against 0.5 s or 0.2 s and now allow 2.0 s, so a loaded runner does not fail them.
3. Locale-stable gates. ci-gates.sh rules and tests-wired run their comm calls in the C locale, like the sort -u calls that feed them. Under en_US.UTF-8 a short baseline could report a baselined rule as new; the shipped baseline is empty, so nothing had failed yet. Both locale tests print a SKIP line and still run the C-locale cases where en_US.UTF-8 is missing.
4. Faster, offline tests. test_auto_update_legacy_marker.py skips docs/bench and runs in about 4 s instead of about 326 s; it and two other tests set DCR_NO_PROBE=1, so none of them calls github.com.
5. Audit PR-body footer. comms_lint --kind audit-pr accepts the attribution footer inside bounded markup (<sub>, <sup>, <small>, *, _); any other wrapped line and anything after a footer are still checked.
6. Scan-summary test. test-review-checks-scan-summary.sh unsets REVIEW_SEMGREP_REGISTRY, so a host that sets it still passes, and a new suite runs it with the variable set.

Verification

• All 41 CI suites this change can affect pass locally, run one at a time through the scrub_env.sh wrapper that run-ci-tests.sh uses, on da58c6b9e: the same changes before the 1.622.0 release stamp was merged under them. 39 passed in the batch run. secret-rule-waves.sh failed there because the worktree lacked the gitignored .banlist.local.txt that the privacy gate requires when the main checkout has one, and security-gate-waves.bg.sh failed once when the wall-clock case of test_destructive_guard_w21.py did not end within its 20 s wait at a 15-minute load average near 13. Both passed when re-run alone, with no code change.
• On this head (fc265cede, rebased onto the 1.622.0 release merge 5ebd7ea8b): the commit-identity gate, the size ratchet against 5ebd7ea8b, the changelog fragment check, git diff --check and tests-wired (480 tests: 468 run by CI, 12 allowlisted) pass. Outside the regenerated SHA256SUMS, its diff has the same git patch-id as the tested da58c6b9e.

test-ci-gates.sh is left to CI. Each lane was independently reviewed, and the wave review's one yellow (sub-second wall-clock pins in newly registered tests) and three blues were fixed before assembly.

🤖 Generated with Claude Code