[Skills] Add bounded autonomous recovery for local lintdiff blockers

[Yuchao Yan (msyyc)]

Why

Follow-up to the merged #5552, based on the completed AvoidAnonymousTypes development #5542 and promotion #5551. The user explicitly requested reducing authorization-only interruptions and avoiding duplicate optimization. This branch starts at b5f23f611, which contains #5552 (69160c487).

The earlier stops obeyed the then-current finite budgets; they were not missing credentials or technical guidance from the user. This PR deliberately adds narrowly bounded, prospective queue authority rather than treating a generic "give access" as unlimited permission.

Intervention evidence	What actually enabled progress	Remaining change beyond #5552
Original local correction 3/3 and first supplemental 1/1 exhausted on named-model spreads with HTTP metadata. The second authorization supplied no new algorithm.	Inspecting the supported effective-model filter and testing metadata inside/outside/both resolved the residual false positive.	Early matrix/API verification and a coordinator-owned local reserve, not unlimited retries or changing source semantics during promotion.
Development review exposed ambient diagnostic changes in other rules' fixtures. Recovery used a normal target merge, corrected CRLF serialization, and reconciled explained snapshots.	Complete shared-fixture coverage and strict serializer-compatible reruns, not new user requirements.	Discover affected fixture groups before publication and preserve the harness writer's exact bytes. Fetching is explicitly distinguished from integration; no new automatic merge authority.
After three recovery attempts, a formatter wrapper expected 11 eligible files although the actual list contained 10. It stopped before Prettier ran.	One authorized command correction derived the correct explicit file list and completed validation.	Derive/verify actual path membership rather than assert a guessed count; eligible correction can consume the shared finite reserve.
Required Core suite had 5 per-test timeouts; its one-worker diagnostic retained 1. The user approved a proposed investigation without supplying a diagnosis.	One baseline and one restored-draft focused run passed under supported threads/one-worker execution; full suite then passed 430 tests with the same 2 skips and 432 IDs, without timeout changes.	One coordinator-controlled baseline-comparison allowance after the original diagnostic, with exact restoration and full-scope passing gate. Original timeout cause remains unproven.

Changes

Five Markdown skill documents only:

• Keep ordinary three-attempt phase/round budgets. Add a three-attempt total per-rule-task reserve, allocated one attempt at a time by the coordinator only for proven in-scope local defects and evidence-based changed approaches.
• Add one native comparison allowance per rule task after a naturally completed, reduced-concurrency diagnostic still reports only per-test timeouts: at most one baseline focused run, one draft focused run, then one original-full-scope run.
• Route both as nonterminal same-owner handoffs; preserve counters across development, promotion, reviews and repair cycles. No fresh worker/review pair to bypass a stop.
• Wire queue, review and promotion instructions to the shared contracts.
• Add early metadata/shared-fixture coverage and deterministic formatter path selection.
• Include eight decision cases covering safe continuation and retained stops.

Not duplicated from #5552

No changes to explicit-target worktree adoption, optional-versus-required gate classification, broad-wrapper execution defaults, Chronus normalization, read-only policy reconciliation or post-run review policy. The optional broad-check limitation in #5551 remains unchanged. This PR does not rerun either completed rule task or modify their source/promotion worktrees.

Boundaries

• These are prospective queue allowances, not retroactive permission for tasks stopped under older contracts or extra allowances for standalone skills.
• No retries of authentication/network operations, review requests, publication or email.
• No assertion weakening, skipped fixtures/tests, timeout increases, dependency reinstall, reset/stash, discarded work or unverified baseline runtime.
• Hook/setup timeouts, crashes, hung/killed commands and unknown operational failures are ineligible.
• Baseline and draft focused passes do not replace the full required suite. Any failed or changed-population full rerun remains blocking.
• Preserve immutable promotion source, independent review and exact publication gates.
• Reserve exhaustion still stops. Generic "continue" does not replenish it.

Validation

Documentation-only validation passed:

• Scoped Prettier write/check for the exact five Markdown files. Reused installed Prettier with the checkout's root style options and no non-Markdown plugins; no dependency or submodule installation.
• All seven added relative Markdown links/anchors resolve.
• Complete diff reviewed, including the eight decision cases, queue/standalone boundaries, explicit user limits, preserved ordinary counters and same-owner continuations.
• git diff --check passed.
• Fresh canonical migration-target fetch still matches the branch base; no intervening skill changes to duplicate.

Decision cases are policy checks, not a claim of an end-to-end agent replay. No package builds or corpus runs are needed for this Markdown-only change.