[finding] plugins/uberdev/skills/premerge-pipeline/SKILL.md — 3 findings #777

Origin: 2c6db44177351a84156c12e838aceecd76f8713b
Agent: premerge.review.code_review
File: plugins/uberdev/skills/premerge-pipeline/SKILL.md:1368
Severity: suggestion (Phase premerge)
Disposition: DEFERRED (premerge-repair-budget-exhausted)
Tier: SUGGESTION
Related: PR #768

Findings (3)

1. plugins/uberdev/skills/premerge-pipeline/SKILL.md:1368 — suggestion / SUGGESTION / DEFERRED (premerge-repair-budget-exhausted) — SKILL.md still says STOP_NO_PROGRESS and STOP_REGRESSED are the ones that matter, and line 39 still lists only those two

Lenses: premerge.review.code_review

The decision table gained the STOP_SELF_REFERENTIAL row and the section below explains it at length, but the two summary sentences that frame the loop's contract were not updated. commands/premerge.md and references/common-mistakes.md WERE updated to name all three, so the pipeline file that agents actually execute is now the one place still describing a two-stop loop.

2. plugins/uberdev/skills/premerge-pipeline/SKILL.md:1414 — suggestion / SUGGESTION / DEFERRED (premerge-repair-budget-exhausted) — Both operator-facing surfaces say the stop needs two consecutive measured rounds at or above the ceiling; the code addit

Lenses: premerge.review.code_review

_growth_round_counts ends `return blockers >= previous_blockers`, and cmd_converge's comment is explicit that this is ELIGIBLE, not merely measured. An operator watching a run print GROWTH=0.90 twice and NOT stop has, from the docs, been told the detector is broken. The inverse is worse: someone simplifying the predicate back to two conditions to match the documented contract re-introduces the measured 5-2-2 false stop the condition was added to fix.

3. plugins/uberdev/skills/premerge-pipeline/SKILL.md:1433 — suggestion / SUGGESTION / DEFERRED (premerge-repair-budget-exhausted) — SKILL.md and common-mistakes.md both enumerate exactly FOUR causes of GROWTH=- and omit the Phase-3 CI-repair arm=commit

Lenses: premerge.review.code_review

_read_repair_scope's docstring states verbatim that a committed-nothing repair is not the only unmeasured case, that the arm=commit path pushes real code edits while publishing only ci-scope-<NN>.*, and that STOP_SELF_REFERENTIAL is therefore unreachable for a CI-repair-heavy loop. SKILL.md lists four causes; common-mistakes.md repeats the same four. An operator on such a run sees GROWTH=- on every attempt, checks the four causes, finds none applies, and concludes the ratio is fine.

To resolve: address the findings above in code and close this issue. Closing resolves exactly the findings listed in this body and the findings added by a later comment on it — one index records both, and closing resolves that recorded set and nothing else. A finding in this file that this issue never recorded is filed as a NEW issue rather than suppressed, so closing never mutes the file. Before closing, apply finding:true-positive if it was a real defect or finding:false-positive if it was not — that label is the eval ground truth (RFC 0018).