fix(store): acquire reindex writer gate unbounded to stop warmup crash

[zzet]

Problem

The daemon panics during warmup:

panic: store_sqlite: store_sqlite: reindex writer gate: context deadline exceeded
  ...store_sqlite.(*Store).ReindexEdges
  ...resolver.(*Resolver).persistAttributionReindexes
  ...resolver.(*Resolver).bindDataflowCalleeRefsForFile
  ...resolver.(*Resolver).ResolveAll
  ...indexer.(*MultiIndexer).RunPreEnrichResolve
  ...main.warmupDaemonState

Root cause

reindexEdgesSetOriented acquired writeMu with a bounded LockContext(sqliteBusyRetryWindow) — 15s by default. A resolver reindex is a mandatory graph.Store mutation with no error channel (the interface mirrors the in-memory store's "everything succeeds" contract), so ReindexEdges routes any failure through panicOnFatal, which only swallows ErrNoRows / ErrConnDone / store-closed — not context.DeadlineExceeded.

When a sibling writer (a warmup WAL checkpoint, or a mutation from the deferred-enrichment pool started ahead of the resolve) held writeMu past the 15s window, the gate timed out and the recoverable lock-wait was escalated into a whole-daemon panic.

reindexEdgesSetOriented was the only mandatory-write path using a bounded gate; the siblings ReindexEdge, setEdgeProvenanceBatchSetOriented, and RemoveEdge all use a plain unbounded writeMu.Lock().

Fix

Acquire the gate with a plain unbounded Lock(), matching the sibling mandatory-write paths. A mandatory graph mutation must wait for the writer gate, not abandon the batch and crash.

Waiting cannot hang:

• beginWriteContext never re-locks writeMu;
• the only unbounded writeMu holder — Compact's VACUUM — runs sequentially before the warmup resolve in the same goroutine, never concurrent;
• lock order is always resolver mutex → writeMu, no inversion (the siblings already prove unbounded Lock is safe from these resolver contexts).

Per-transaction SQLITE_BUSY contention remains bounded by withSQLiteBusyRetry. The checkpoint's bounded LockContext is left untouched — a checkpoint is retryable maintenance whose caller treats failure as skip-and-continue.

Verification

• go build ./... and go vet ./internal/graph/store_sqlite/ clean.
• store_sqlite busy / reindex / checkpoint / provenance tests pass with -race and without: TestCheckpointSerializesBehindWriterGate, TestReindexEdgesRetriesWholeTransactionAfterBusy, TestReindexEdgesPersistentBusySurfacesAndRollsBack, TestLongWALReaderDoesNotBlockReindexWriter. No test asserted the gate-timeout path, so behavior is unchanged for covered cases.