P0: Commit a Stop Condition and Round Budget for the Whole-Unit and Local Review Passes #1312

Part of #1311. Priority: now. One pull request.

Rescoped 2026-09-04 after the first attempt, parked unmerged on feature/review-stop-condition and assessed in this #1311 comment and the re-assessment that follows it. That branch wrote the stop rule as roughly 1,200 words of new contract prose across two units and did not converge in twelve rounds. It is superseded by the scope below, is not merged or rebased, and is deleted once this issue's pull request merges.

Why

The whole-unit canonical pass and the local strict review pass have no committed stop condition, so the loop runs until an agent decides on its own to stop, and #1267 measured that it does not converge: PR #1198 ran eight fix-and-re-read rounds where nearly every round's findings were introduced by the previous round's fix. The rule's home is local-strict-review "Disposing of Findings", which is one fourteen-sentence paragraph carrying 40 of #1327's 77 pre-existing findings, so landing the rule beside that paragraph leaves in place the text it contradicts. This issue is therefore also #1311's deletion pilot: the paragraph is rewritten by deletion into a normative block, with the stop rule inside it, and the rewrite is measured.

Scope

• local-strict-review "Disposing of Findings": replace the paragraph with a short normative block. Each rule is one bold statement of at most two sentences, followed by a marked non-normative "Why:" line. Content, and nothing beyond it:
  • Findings map to pr-review-conduct's five outcomes by pointer, not re-enumerated.
  • Each finding carries a class. introduced: the change wrote, rewrote, or removed the text, or the text should have been written, plus the two carve-outs from The whole-unit review loop has no committed stop condition, so it does not converge #1267, a finding load-bearing for a maintainer decision this change puts, and one showing the change's own edit left a precondition elsewhere false. pre-existing: anything else on text the change did not write. style: a preference between two defensible wordings.
  • The pass closes on one question: does content this change wrote still carry a false or contradictory claim, an instruction that fails when followed literally, or wrong behavior. Introduced is fixed. Pre-existing is filed once and blocks nothing. Style on text already rewritten twice takes the shortest defensible wording and stops.
  • The budget: two rounds of edits per push over the same content, then the remainder goes to the maintainer with its class counts.
  • One sentence keeping the pass mandatory and its findings advisory.
• local-strict-review "Running It": the Return: line gains the class, and the brief gains one line: for every rule the diff adds or changes, grep the canonical tree for other statements of it and report each that now disagrees.
• GOVERNANCE.md "Verification Discipline": one bullet of at most three sentences. The stop question is the rule. The class, the bound, and the report line are the Skill's, and the bullet does not restate them.
• pr-review-conduct "Never exit the loop early": one clause scoping that rule to the PR-hosted loop and naming local-strict-review as the home of the pre-push bound.
• backlog-burndown: not touched. Its shape is P2 (decision): Decide backlog-burndown's Shape Before Its 23 Open Defects Are Worked One at a Time #1323's decision, and the pointer replacement for its "set a review-round budget" bullet is filed there.

Rules that bind this task

• The rule text lives in GOVERNANCE.md and the procedure in the Skill. Neither restates the other.
• The pass stays mandatory and its findings stay advisory. This task changes when the loop ends, not whether it runs.
• The pilot runs under the rule it writes, from the first round: the stop question and the two-round budget go into every reviewer brief at spawn.
• The measurement is at most six passes: three whole reads of the rewritten unit plus one diff pass carrying the sweep line, each record carrying --findings, findings classed per round. If the unit has not converged inside that budget, stop, and that is the result.
• Regenerate the skill mirrors with python3 scripts/build_dist.py, never by hand.

Acceptance

• An agent following local-strict-review on a branch whose second round reports only pre-existing and style findings pushes without a third round, and the report says why.
• The rewritten unit's finding curve over its three whole reads is posted on Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311, converged or not.
• The Pre-existing defects in the five review-loop canonical units, found by the #1312 passes #1327 findings against "Disposing of Findings" are each closed as settled by the rewrite or moved to that unit's tracker, with the counts stated.
• The whole-unit review loop has no committed stop condition, so it does not converge #1267 closes. Local code review should be executed before every PR review #1104 is answered by the existing every-push rule plus this budget, and closes or is re-scoped.

Closes #1267. Relates #1104, #1283, #1327.