[TheAmericanMaker]

Summary

Item 3a of the 2026-09-14 recommendation, with the number the release was held for.

The verify pass (#331) took the 2026-09-13 sweep's top twelve from 17% precision to 12/12 agreement with a reviewer by applying one rubric: a finding is a reachable failure — name the trigger; a cast every caller satisfies is not a defect. This PR puts that rubric into the defect lens's own batch prompt, after the per-language pattern list, so the noise is cut where it is produced rather than paid for twice. Type hygiene that is worth a note goes at severity low under the pattern name type-hygiene, and the reading guide says to read those rows as notes.

Measurement

Same repository (this one, at bcb34ba), same model (google/gemini-3.7-flash:batch, low effort), defect lens only, two runs per arm. Every finding at medium was judged by hand against the source before reading the verifier's verdicts; lows were judged on the rubric runs and on the second baseline run.

arm	run	findings (med / low)	true at medium	medium precision	true overall	cost
baseline (0.24.0 prompt)	2026-09-13T20-00-19-293Z	24 (12 / 12)	amendment lock, --namespace	2/12 = 17%	2	$0.102
baseline	2026-09-14T04-05-15-830Z	26 (13 / 13)	amendment lock, --namespace, linger timer	3/13 = 23%	3	$0.104
rubric (this PR)	2026-09-14T03-53-20-895Z	10 (4 / 6)	amendment lock	1/4 = 25%	1	$0.097
rubric	2026-09-14T04-05-15-533Z	13 (6 / 7)	amendment lock, --namespace, stale-lock break race	3/6 = 50%	4 (+ renderSafeLink at low)	$0.096

• Volume halves (mean 25 → 11.5); medium precision roughly doubles (mean 20% → 38%); real defects per run are the same on average (2.5 → 2.5), and either prompt misses a real one on some runs — the batch pass is a sample, which is what verify and a second run are for.
• verify on the two second-sample runs: 11/12 agreement with the hand judgment on each (each "confirmed" one hypothetical the code does not reach: a pre-aborted signal no caller passes; a ./-prefixed primary_output). $0.14 and $0.16.
• Four real defects surfaced across the four runs: amendment lock (fix: amendments judge pipeline completeness on the state read under the lock #337, fixed), --namespace (fix: /codecarto-library-init refuses an incomplete or invalid --namespace #336, fixed), the clearPhase linger timer on re-entry, and the stale-lock break race in acquireLock — the last two are filed as issues.
• n = 2 per arm on one repository. The direction is consistent in both pairs; the magnitude is not a measurement.

Recommendation: merge — it is free, it halves what a reader has to dismiss, and it does not cost recall on this evidence.

Test plan

• tests/broadside.test.mjs: the rubric is pinned in the defect prompt for every language profile, after the pattern list.
• npm run build + npm test — 907 pass.
• Live: the two rubric runs above.

🤖 Generated with Claude Code