Add pull request attention queue skill

[PureWeen]

Goal

GitHub notifications are no longer a useful intake queue for a team this size. We have framework work, community contributions, vendor PRs, and broad team review requests all mixed together. The result is that useful PRs can sit unnoticed, while a notification by itself does not tell us whether someone can review the PR now or whether it first needs ownership, author follow-up, CI work, or a disposition decision.

The longer-term goal is an ASP.NET Core team canvas that gives us a small, explainable daily view of the work that deserves attention. The first canvas is in the stacked PR #69063.

This PR is the starting point: a reusable, read-only skill that owns the deterministic classification and ranking logic. Keeping that logic outside the UI gives us something the team can inspect, evaluate, and iterate on together before we expand the canvas or add other surfaces such as issue triage.

What this skill does

Adds .github/skills/pr-attention-queue, which produces an actionable pull request attention queue.

• Separates Review now, where a human reviewer can productively act, from Needs rescue, where a maintainer or triager must restore ownership or decide how the PR should proceed.
• Distinguishes never-reviewed work from reviewed work where the author responded but reviewer follow-up was abandoned.
• Identifies ready-to-merge and other waiting states without mixing them into the daily review queue.
• Supports named presets and ad hoc label/path filters, defaulting to a Blazor preset.
• Reports the resolved scope, next actor, blockers, and evidence-backed reason codes instead of producing an opaque AI priority score.
• Emits Markdown or versioned JSON with stable display metadata for canvas, workflow, or other consumers.
• Performs no GitHub mutations.

The initial ranking is deliberately opinionated but not intended to be final. The useful feedback on this PR is whether the buckets, ordering, thresholds, scope rules, and next-actor decisions match how the team actually handles incoming work.

Correctness hardening from the live pilot

A complete live Blazor pilot found several places where high recall was not yet precise enough for unattended recommendations. This revision keeps the same deterministic model while tightening those boundaries:

• recognizes the repository's exact * NO MERGE * veto and pending-ci-rerun signal with explicit precedence;
• requires GitHub mergeStateStatus == CLEAN before a PR can enter Ready to merge;
• routes an approved BEHIND PR to author/maintainer branch-update work instead of attributing it to CI;
• excludes author-authored review records from reviewer evidence;
• lets current non-author reviewer feedback override an older outstanding team request, while preserving the reviewer roundtrip when the author later responds or pushes;
• emits a one-based digestRank so JSON, Markdown, and canvas consumers render the selected order consistently;
• supports explicit digest-only author exclusions without removing those PRs from the matched universe or bucket census;
• detects bounded, same-repository stack ancestry and prevents a reviewable child with unhealthy ancestors from consuming an unattended digest position without rewriting its underlying bucket;
• ignores cross-repository fork branch names when resolving stack ancestry, avoiding false links to common branch names such as main.

These are conservative gates. False negatives are preferable to telling a merger that blocked or unstable work is ready.

Canvas direction

#69063 is stacked on this PR and shows the intended product direction. It uses this skill as the canonical engine and adds an Aspire-style canvas with:

• separate Review now and Needs rescue lanes;
• a compact Ready to merge view and expandable secondary classifications;
• explicit reasons, blockers, age, scope, and next actor;
• narrow read-only actions to open a PR, start a review session, or investigate rescue work.

The canvas does not duplicate or override the skill's classifications. The idea is to let us improve the shared skill based on real team feedback while the UI remains a thin consumer of its versioned output.

Implementation and validation

The skill uses a thin PowerShell entry point (scripts/Get-PRAttentionQueue.ps1) over scripts/PRAttentionQueue.psm1, consistent with the existing PowerShell scripts under eng/scripts.

It includes fixture-based tests (tests/Test-PRAttentionQueue.ps1, with no Pester dependency) and a repository eval specification at eng/skill-evals/pr-attention-queue/eval.vally.yaml.

Validation after the correctness pass:

• all deterministic fixture and contract tests pass, including the live failure modes, review-precedence transitions, fork-branch collision, BEHIND routing, and digest controls;
• a complete live Blazor query reconciled all 345 open PRs and classified 137 matching PRs with no unresolved mergeability;
• Unescape relative path before file lookup to fix space-in-filename issue #62349 moved from false-ready to Needs rescue because of * NO MERGE *;
• [10.0] [Blazor] Support NavigateTo when enhanced nav is disabled (#52267) #64724 moved from false-ready to Waiting on CI because of pending-ci-rerun;
• Fix BL0012 false positive on await foreach and await using #68720 moved from false-ready to Waiting on CI because its merge state is UNSTABLE rather than CLEAN;
• ErrorBoundary fails to show error content when exception occurs inside @foreach #67177 is now Waiting on author, while Fix null reference issue caused by validation exception when circuit is disposed #67146 remains Review now because its author responded after review;
• [Blazor] Generate AOT-safe binding metadata #68297 keeps its reviewable classification but is excluded from the digest while unhealthy stack ancestors remain;
• explicit exclusion of PureWeen retained those PRs in the census while removing them from the visible digest;
• Markdown and the stacked canvas render the same engine-selected rank order;
• an independent final code review found no remaining significant issues after the stack, merge-state, and forwarding corrections.

This is a draft so we can validate the model with the team before treating the current ranking policy as established behavior.