fix(bindx): announce the writes that made the store go stale

[matej21]

Stacked on #80. Its base is test/store-notification-pins, not main — #80 adds the reproductions, this removes them. Merge #80 first, or say the word and I will squash the pair into a single PR against main.

Two SnapshotStore write paths mutated observable state without notifying. #80 reproduced both with test.failing pins; this fixes them, which turns those pins red (this test is marked as failing but it passed) and so also un-marks them into ordinary regression tests.

createEntity — the user-visible one

this.setEntityData(entityType, id, data, false)   // notifies
this.setExistsOnServer(entityType, id, false)     // notifies
this.roots.register(this.getEntityKey(entityType, id))   // silent

The root registration is the write that makes the entity count as a create — the comment directly above it says so. Both notifications therefore carried the pre-registration value.

Observable effect: mount a plain <Entity create> and a save indicator built on store.subscribe + getAllDirtyEntities().length — the shape usePersist uses — renders 0, while getAllDirtyEntities() returns one create.

Reordered, not re-notified

setEntityData  ->  roots.register  ->  setExistsOnServer

Adding a notification after roots.register would have worked too, and was rejected: it makes useEntityList adding N drafts cost 3N notifications instead of 2N. The reorder costs nothing.

The position is forced from both sides:

• After setEntityData — ReachabilityAnalyzer.walk() seeds a root only if (entitySnapshots.has(key)), so registering first would be inert, and would leave a dangling root if the snapshot write threw.
• Before setExistsOnServer, which then carries the final notification. Moving it last is observationally free: EntityMetaStore.existsOnServer defaults an unknown key to false, so for a fresh entity that call is bookkeeping.

roots.register bumps mutationVersion, invalidating the reachability memo, so the subscriber woken by that last notification recomputes and sees the create.

Undo journal is unaffected: the pre-image is captured inside setEntityData (journal?.recordEntity(key)) before any root write in either order, so exportEntityCell's isRoot / present:false capture is identical. The per-kind write guard fuses snapshot + meta + roots into one entity kind, already recorded by that same call.

clearAllServerErrors

Silent, where its sibling clearAllErrors notifies for the same kind of write, and getVersion() did not move either — so no consumer, entity or global, could observe the clear.

Both in-tree callers self-heal by accident of ordering (BatchPersister dispatches a notifying setPersisting(true) right before; mapServerErrors follows every clear with a notifying addFieldError), so this only ever reached users through the exported action. Their re-render count goes from 1 to 2 per persist attempt — not a per-keystroke path.

Not fixed, recorded

Both siblings clear relation errors under the key prefix but notify only the entity, so a subscribeToRelation consumer still sees a stale value after either call. Fixing that asymmetry would change the notification counts these tests now pin, so it is a separate decision.

The two other suspects from #80 stay untouched and their characterization tests are unchanged: sweepUnreachableCreated delegates to a notifying removeEntity, and unregisterRootEntity's callers sweep on the next line. Neither is a bug.

Gates

• bun test tests/unit/store/ tests/react/storeNotifications/ — 194 pass / 0 fail
• bun run typecheck — exit 0
• bun test --path-ignore-patterns='**/tests/browser/**' — 1751 pass / 0 fail / 154 files, identical totals before and after (the delta is only the 5 pins flipping red→green)

Negative control: with the store fix reverse-applied and the markers still off, exactly the 5 un-marked tests fail. The two control and two characterization tests pass in both directions.

Comment headers in the four files were rewritten from "known-broken pin" to regression notes — they described test.failing semantics that no longer apply. No assertion, test name or expect was touched; verified by diffing non-comment lines and by the unchanged expect() count (1959).