[Bugfix #844] Gate NeedsAttention PR rows on porch pr-gate

[waleedkadous]

Summary

Fixes #844

Root Cause

packages/dashboard/src/components/NeedsAttentionList.tsx:42-56 iterated every open PR from overview.pendingPRs and emitted an attention row unconditionally. There was no cross-reference against the builder's porch gate state, so a PR appeared in Needs Attention the moment it was created — before the automated CMAP reviews (Gemini / Codex / Claude) had run. Reviewers arrived ahead of the AI review comments.

Fix

buildItems now consults OverviewBuilder.blocked / blockedSince to decide whether a PR is genuinely waiting on a human:

1. Build prGateSince = Map<issueId, blockedSince> for builders with blocked === 'PR review' (the human-readable label for the porch pr gate, mapped in overview.ts:GATE_LABELS).
2. For each PR, include it iff EITHER:
   • its linkedIssue matches a pr-gate builder (prGateSince.has(...)), OR
   • it has no associated builder AND reviewStatus === 'REVIEW_REQUIRED' (the human-authored / externally opened case — no porch gate to wait on, so fall back to GitHub's reviewDecision).
3. waitingSince uses the builder's blockedSince for affiliated PRs (the "human became bottleneck" moment), falling back to pr.createdAt for unaffiliated PRs.
4. Track emittedPrGateIssueIds; the builder loop only dedupes pr-gate builders whose PR was actually emitted. This protects against a stuck builder disappearing if its PR is missing from prs (cache miss, pagination, transient API failure).
5. gateKindClass gains a 'PR review' → 'attention-kind--pr' case so the builder-loop fallback row from (4) renders with the same PR styling.

The linkedIssue plumbing already existed on OverviewPR (populated by parseLinkedIssue in overview.ts); no API/types change was needed.

Note on issue suggestion

The issue suggested dropping the continue at line 62, but doing so would create duplicate rows (one from the PR loop, one from the builder loop) for the same builder. Kept the skip and conditioned it on the emitted-set so a missing PR doesn't silently hide the gate signal.

Test Plan

• 9 unit tests covering acceptance criteria + waitingSince + missing-PR fallback + dedupe
• Local tests pass: pnpm test NeedsAttentionList → 9/9 passed
• porch check passes (build + tests)

Test scenarios covered

• PR with builder still in CMAP (blocked === null) → excluded ✓
• PR with builder at pr gate → included ✓
• PR with builder at pr gate → emitted exactly once (dedupe holds) ✓
• Human-authored PR (no builder) + REVIEW_REQUIRED → included ✓
• Human-authored PR (no builder) + APPROVED → excluded ✓
• Other gate-pending builders (spec/plan/dev review) still surface ✓
• PR with stale linkedIssue (no matching builder) treated as unaffiliated ✓
• Affiliated PR waitingSince === builder.blockedSince (not pr.createdAt) ✓
• Builder at pr-gate with missing PR still surfaces via builder loop ✓
• Dedupe still holds when both PR and builder are present ✓

CMAP Review (PR-845, 3-way)

Model	Verdict	Confidence	Notes
Codex	APPROVE	HIGH	"Focused, addresses root cause, solid regression coverage."
Claude	APPROVE	HIGH	"Clean, well-scoped." Flagged waitingSince as a non-blocking observation.
Gemini	REQUEST_CHANGES	HIGH	Three points: waitingSince semantics, invisible stuck builders, missing gateKindClass case.

All three of Gemini's points addressed in iter-2 commit 7ec0fc9c.

[waleedkadous]

Excellent work — thank you. Approving.

Two things I particularly liked:

1. Pushback on dropping the continue — the issue suggested removing it, you kept it and made it conditional on emittedPrGateIssueIds.has(issueId). That's a strictly better solution than what I proposed: it preserves the dedupe and gracefully handles the cache-miss case (PR not in prs yet, builder still surfaces). Good judgment.

2. waitingSince = blockedSince, not createdAt (Gemini's catch). This is the right semantics — the waiting-time chip now measures "human-as-bottleneck" duration, which is what the Needs Attention surface is for.

The regression test for the double-emission invariant is the kind of test I'd hope someone would write but rarely see. Nice.

Please merge.