Late comment reads can erase newer comment mutations from the shared cache

[Chris0Jeky]

Summary

cardCommentStore.fetchCardComments commits every response directly into cardCommentsByCardId, while create/update/delete mutate the same cache independently. The modal's generation token protects only its local loading/error state; it does not prevent the store from committing a stale payload.

Confirmed races

1. A comment fetch begins.
2. A create, update, or delete succeeds and updates the shared cache.
3. The older fetch resolves with its pre-mutation snapshot and replaces the cache, erasing the confirmed mutation from the UI.

Two overlapping fetches can also resolve out of order, and a request that outlives board navigation can repopulate the cache for the board that is no longer selected.

Expected contract

• Only the latest read for a card may commit.
• A read that began before a successful local comment mutation may return its payload to its caller, but must not overwrite the post-mutation cache.
• Late requests must not repopulate a different selected board's comment cache.
• Create remains idempotent by stable comment id when an authoritative refresh committed the same comment first.
• Existing API return values, sorting, toast/error/loading behavior, and demo-mode behavior remain unchanged.

Acceptance

• Focused regressions for stale fetch after create/update/delete.
• Regression for out-of-order reads.
• Regression for navigation while a read is in flight.
• Regression for refresh-before-create stable-id deduplication.
• Targeted frontend tests, typecheck/build, and exact-head hosted CI pass.

[Chris0Jeky]

[Codex lane claim v2]

• lane: alpha-product-trust
• base: 307c3b8b50bec1cb0bfaea3e570a942bcb1d4451
• owned paths: frontend/taskdeck-web/src/store/board/cardCommentStore.ts plus one focused concurrency spec
• shared-path leases: none
• parallel-safe work: no overlap with the active card-delete or label-ownership branches
• status: pin stale-read-after-write, out-of-order read, navigation, and stable-id dedupe regressions; then add the smallest per-card cache ownership/version seam and validate the exact head

[Chris0Jeky]

Claiming the bounded integration of #3303 / source PR #3304 for the current Codex coordinator. Fresh inventory confirms source head 66e725d is ready, has no active writer/worktree, and all five existing review threads are resolved. Its completed review rounds remain closed; this is an integration check against the newer session and loading contracts, not a new source audit.

The isolated integration will preserve the source commits and three regression files, reconcile cardCommentStore with #3358 and #3359, and prove queued transport, late settlements, cache ordering and loading ownership together. Parents merge first. Active implementation budget: about 30 minutes, with a factual parked handoff if unresolved blockers consume roughly twice that.