Merge-readiness poll gate write can be clobbered by a concurrent hydration wave's whole-map republish

[jaylann]

Summary

publishChecks (gbar/Sources/Store/AppStoreHydration.swift) republishes the hydration maps as whole-map assignments (prGates = pending.gates), where pending.gates was snapshotted from prGates at drain start. If startMergeReadinessPoll's refreshPRState writes a freshly-unblocked gate under the same checksGeneration between that snapshot and the publish, the wave's wholesale reassign overwrites it — re-hiding the Merge button until the next full refetch.

Pre-existing (the old code already did prGates = pendingGates from a snapshot); noting it now because the .checksOnly path added in #82 widens the window — a skipped PR's gate in the snapshot is stale by construction rather than freshly fetched.

Repro (timing-dependent)

1. Viewer approves PR refactor(design): deeper, richer state colors #7; mergeable_state still async-recomputing, so startMergeReadinessPoll begins polling (captures generation = G).
2. A regular 60s poll fires and starts a hydration wave. If that wave reads checksGeneration == G (i.e. it didn't bump generation before the merge poll captured it, or they interleave under the same G), drainChecks snapshots pending.gates = prGates — with refactor(design): deeper, richer state colors #7's old "blocked" gate.
3. The merge poll's refreshPRState writes prGates[#7] = the newly-unblocked gate (Merge should show).
4. The wave completes and publishChecks does prGates = pending.gates, clobbering refactor(design): deeper, richer state colors #7 back to the stale "blocked" gate → Merge disappears until the next full refetch re-derives it.

Both run on @MainActor but interleave across await suspension points. The checksGeneration guard prevents cross-generation clobbers; the gap is a same-generation interleave.

Options

• Merge refreshPRState's single-key gate write into the wave's pending map rather than writing prGates directly (or have the wave re-read per-key gates at publish time instead of a snapshot).
• Have refreshPRState bump/own a generation so a concurrent wave's publish is dropped for keys it touched.
• Per-key gate publishing for the merge-poll path instead of whole-map reassign.

Low severity / narrow timing window and self-heals on the next full refetch, but it can briefly hide a valid Merge button right after an approval — the exact UX the merge-readiness poll exists to fix.

Surfaced in the #82 code review.