docs: add re-scope-after-a-regression rule to task-writing.md

[bborbe]

What

Adds a rule to docs/task-writing.md § Audit: re-scope after a regression, never patch.

A criterion that has reached a passing verdict and then regresses is re-scoped or cut, not repaired again. The trigger is the first regression after a held verdict; the action is to cut the criterion back to what the evidence can carry, or replace the claim with one the artifact supports — never to add a clause narrowing the failing one.

Why

Measured on two independent task rows in one manager session (2026-09-26) that cycled 14 and 5 audit rounds without converging. Neither reached the 9/10 bar, and both were ended by a re-scope, not by a successful patch.

The rejected alternatives are recorded in the section, each with the failure it would produce:

• A clause budget — satisfiable by deletion, so it measures length rather than convergence. The fatal defect is typically a single clause inside an otherwise reasonable criterion, so a budget flags it and then watches it cycle.
• Delete at least as many clauses as you add — bounds the growth rate rather than the duration, and a round that changes nothing satisfies it while making no progress.
• Re-scope after N rounds — no value of N works. N=3 discards a cycle that reached 8/10 by patching; N=5 still misses the actual failure. Both rows improved by patching before they broke by patching, so a count cannot separate the productive rounds from the destructive ones.

Hence a regression after a held verdict rather than a count. Checked against both rows before shipping: it fires at round 10 of 14 in the first — four rounds before the re-scope that actually ended it — and at the last round of 5 in the second, where it ratifies rather than accelerates.

Verification

make precommit green — all 8 Go packages ok, scenario suites 0 failed, golangci-lint 0 issues, CHANGELOG structure OK.

[ben-s-pull-request-reviewer]

The review is complete. This is a clean docs-only PR with no issues.

Summary: A docs-only PR adding a well-reasoned "re-scope after a regression, never patch" rule to task-writing.md § Audit. The section is correctly placed as a sub-section under Audit, grounded in empirical evidence (two observed cycling rows of 14 and 5 rounds), and clearly distinguishes the rule from rejected alternatives (clause budget, delete-symmetry). Version alignment is verified across all four required files.

{
  "verdict": "approve",
  "summary": "Docs-only PR adds a well-integrated 're-scope after a regression' rule to task-writing.md § Audit, with empirical grounding and clear rejection of alternatives. Version alignment verified across all four required files. No concerns.",
  "comments": [],
  "concerns_addressed": []
}