Access metadata writes

Go storage providers implement WriteAccessMeta(ctx context.Context, request AccessMetaWriteRequest) (written bool, err error). Callers must choose an explicit, nonzero mode and pass the lifetime of the operation that owns the write. The previous PutAccessMeta entry point is retired; implementations and plugins must rebuild.

ctx owns admission, buffered-order drain and presence preparation. Once the publishing update is admitted the write and its notification complete in the store, so a canceled context returns its own error and leaves no published entry. The access flusher passes its own flush-cycle lifetime: it has already drained those deltas from the accumulator, and the run lifetime is canceled before its final flush, so a request or run context would abandon accepted counters. A store that is reached through a context-free queue entry (PutDeindexWorkItem, DeleteDeindexWorkItem) passes that entry's own explicit unbounded lifetime instead.

Mode	Contract
knowledgepolicy.RecordState	Record or restore state without requiring a live entity. Preserve internal work keys and the existing index-catalog merge.
knowledgepolicy.FlushExistingEntity	Write only if the stored node or relationship exists at the atomic write boundary. Hidden entities remain eligible. Return false, nil for absence and preserve any retained state.

written, err := engine.WriteAccessMeta(ctx, knowledgepolicy.AccessMetaWriteRequest{
    Mode: knowledgepolicy.RecordState,
    EntityID: id,
    Entry: restoredEntry,
})

The native conditional writer uses the existing entity authority for both node and relationship IDs, orders after acknowledged buffered writes, and checks the stored-inclusive entity in the same managed transaction that writes the metadata. A lookup earlier in the flush is insufficient: deletion may commit after it. No new per-entity cache or lock registry is added.

The flusher skips genuine absence and retains real lookup/write errors for shutdown reporting. GetEntityMeta distinguishes absence with knowledgepolicy.ErrEntityMetaNotFound, also preserving errors.Is(err, storage.ErrNotFound) in the native provider. A node decoding or storage error is propagated rather than retried as a relationship lookup. Suppression and embedding callbacks run only after an accepted write.

Deletion preserves previously saved access state, and recreated IDs retain the existing ID-based state semantics. No entity-generation rule, Badger key, persisted payload or storage-version migration changes.