[Skills] Prevent avoidable lintdiff workflow blockers

[Yuchao Yan (msyyc)]

Why

The end-to-end EvenSegmentedPathForPutOperation run (#5528 development, #5550 promotion) repeatedly stopped even though the user supplied no rule-design or implementation guidance. The interventions were work-reuse authorization, a retry request, one precisely scoped single-worker retry authorization, and a later end-to-end continuation request.

This review traced those stops through the prior session, append-only execution log, preserved validation manifests and current skill instructions. It does not treat all stops as spurious or infer that generic access overrides an explicit validation condition.

Observed evidence	Root cause	Correction
Initial CLI invocation found clean, pushed task-owned commit efbc4cdff, canonical push permission, and no PR, but stopped for legacy-worktree adoption. The earlier environment had refused publication before a creation call because its app publisher lacked the required base controls.	App-specific ownership adoption leaked into general preparation rules; an unattempted operation was treated like a failed publication retry.	Backend-scoped, verified explicit-target resumption. Preserve ownership, state, history and exact-PR reconciliation; no separate app adoption for ordinary CLI reuse.
Required promotion validation passed: 12 focused tests, 486 ARM tests, five ruleset tests, builds/lint/docs/format/Chronus and independent local review. The optional broad wrapper then failed with ten timeouts and was stopped at its deadline.	Promotion called the broad wrapper optional/best effort, but also included it in default command blocks and said to run it before creation. Generic stop rules and inconsistent Validation blocker wording overrode its intended gate level.	Predeclare required versus supplemental gates, separate the optional command from required examples, and disclose scoped supplemental limitations without automatically stopping publication.
A user-authorized --maxWorkers=1 retry took 1178.59 seconds and still failed: 3466 passed, two Autorest setup-hook timeouts, four skips. Later publication used exactly the same ten promotion files, not a code fix or another successful test run.	The agent pursued full-suite success before deciding whether that suite was necessary. The next user intervention supplied no technical diagnosis.	Read-only coordinator reconciliation before terminal stops or requests for renewed permission. Never infer an environmental cause, silently promote a failed required gate to optional, or treat a retry allowance as unlimited recovery.
The promotion correction budget was consumed by an ESLint probe in an oxlint checkout, a diagnostic-marker correction, and Chronus LF normalization.	Preventable command assumptions and Windows line-ending work mixed with genuine draft correction.	Derive tool probes from the target checkout. Normalize named Chronus files before validation, without staging. Keep existing finite correction budgets; the diagnostic-marker fix remains ordinary local correction.
The successful final run reported no qualifying process improvement despite the earlier authorization-only stops.	Post-run review focused on the successful final state rather than the transitions requiring human intervention.	Review each stop, the actual human input, and what enabled continuation; success does not erase a proven instruction contradiction.

The Autorest timeout root cause remains unknown. This PR does not claim baseline reproduction, fix Autorest, rerun the completed rule task, or change its PRs.

Changes

Seven skill documents only:

• Shared recovery contract defines gate/disposition rules, a single read-only policy-reconciliation handoff, and eight decision cases.
• Queue/preparation/publication contracts consistently distinguish explicit-target reuse from app-session adoption and retain the current phase owner/history.
• Promotion keeps its required native checks, removes optional broad runs from default command blocks, and separates Validation limitations from actual blockers.
• Review/fix gates consume that same classification rather than reclassifying every nonzero exit as mandatory.
• Post-run review examines the full intervention history.

Safeguards retained

• Required validation still blocks, including an explicitly user-required broad check. “Publish only if this passes” cannot be undone by a later generic “continue” or “give access.”
• Any plausible task regression remains blocking until classified, even if a supplemental command discovered it. Required checks inside a supplemental wrapper remain required; contradictory failures cannot be hidden by earlier passes.
• No timeout increases, changed assertions, skips, new retry budgets, automatic hook-timeout retries or unrelated fixes.
• No bypass of app-bound publishers, canonical repository policy, folder containment, clean-head/provenance checks, immutable-source repair, or independent review.
• Failed/indeterminate publication requests are not resent merely because an absence query is empty. Historical preflight refusals before any creation call are distinguished from API failures.
• No replacement worktrees, discarded edits, reset budgets or concurrent owners.

Validation

• Full seven-file diff reviewed against freshly fetched Azure/typespec-azure:feature/lintdiff-migration-new at e3fcd29b245a5393beb448872f5018ddc8147c84.
• Scoped Prettier write/check passed for the seven Markdown files. Initial use of the root config failed before changes because its Astro plugin was unavailable in the isolated dependency-free worktree. The successful run reused installed Prettier 3.9.5 with the same root style options and no non-Markdown plugins; no dependencies or submodules were installed.
• Six newly added relative Markdown links/anchors resolve.
• git diff --check passed.
• The documented Windows normalization snippet was smoke-tested on an isolated CRLF fixture: exact content retained, LF output, front matter at byte zero, no BOM, no staging. The fixture was removed.
• Reviewed the eight documented decision cases for safe continuation versus retained stops. These are policy checks, not a claim that an end-to-end agent replay was executed.
• No package, corpus or full-suite runs: this is a Markdown-only workflow correction.

The primary rule worktrees and the coordinator's existing branch are unchanged.