Storage preparation contexts

SchemaManager.GetConstraints, HasAnyConstraintContract, and GetConstraintContractsForTarget take a context and return an error. They use that context for metadata lock admission and enumeration; contract discovery also uses it while cloning entries and sorting. Their only errors are cancellation or deadline errors. A failed enumeration returns no partial result.

constraints, err := schema.GetConstraints(ctx)
if err != nil {
    return err
}
for _, constraint := range constraints {
    // Inspect the detached unique-constraint metadata.
    _ = constraint
}

hasContracts, err := schema.HasAnyConstraintContract(ctx)
if err != nil {
    return err
}
if hasContracts {
    contracts, err := schema.GetConstraintContractsForTarget(ctx, storage.ConstraintEntityNode, "Doc")
    if err != nil {
        return err
    }
    _ = contracts
}

Pass the actual caller context for request-owned preparation. Native CompositeEngine.GetSchema() remains a context-free Engine method: its constituent metadata reads explicitly use context.Background(), for which these cancellation-only errors cannot occur.

PrepareEntityCommit remains caller-owned until its immutable prepared handoff and submission decision. CommitContext remains caller-owned through its final publication decision. Conflict classification, temporal buffering, constraint adjacency, change-image capture, observer preparation, publication dependency keys, and change-event key discovery use the current phase's lifetime. Errors before publication abort the transaction and do not invoke publication callbacks. Pass the explicit preparation context to native point reads, staging APIs and request operation lists as well.

After a prepared command is accepted, its immutable image belongs to the applying engine's accepted-work owner. Caller cancellation can stop receipt waiting but cannot retract the accepted command or turn it into a retryable rejection. Engine shutdown may leave accepted work recovery-pending; its record remains available for recovery. This completion phase does not reuse the original request lifetime or a nil-context fallback. No storage-format change is required.

UNIQUE cache preparation measurement

Cache staging reuses the selected UniqueConstraint claim owner directly, avoiding a temporary SchemaManager for each staged claim. It retains the same typed label/property identity and cancellation contract.

A paired component comparison on Apple M2 Max, Darwin arm64, 12 Go processors, used 10,000 iterations per sample in AB/BA/AB order. Both sides included the same preparation, creation-validation and typed-identity correctness fixes; only the shared claim-owner implementation differed. Containers and local Qwen were stopped during the accepted window; other Go activity was monitored.

Workload	Reference median	Shared owner median	Memory and allocations
Rebuild four UNIQUE caches from 128 stored nodes	161,337 ns/op	139,874 ns/op	185,610 → 185,193 B/op; 1,998 → 1,997 allocs/op
Empty target namespace with another populated database	5,932 ns/op	5,913 ns/op (within observed timing noise)	4,802 → 4,401 B/op; 75 → 72 allocs/op

Run the native populated component with:

go test ./pkg/storage -run '^$' -bench '^BenchmarkSchemaRefreshNativePopulated$' -benchmem -benchtime=10000x -count=3 -cpu=12

The committed benchmark also checks every resulting claim outside the timed loop. These measurements qualify this cache-owner change, not total transaction preparation or end-to-end ingestion. Buffered preparation has a separate workload.

Buffered acknowledged-generation comparison

The matched public preparation fixture uses real acknowledged peer commits after BEGIN, with identical typed-identity prerequisites on both sides. Five alternating fixed30x pairs at12P compare the six-file overlay-admission and poll-amortization correction. The1-node median is29471→27893ns/op with9299B and166allocs unchanged; the64-node median is457968→447690ns/op,435886→435888B and6971allocs unchanged. Timing ranges overlap, supporting no material regression for these fixtures rather than a precise speedup or whole-ingest claim.

BenchmarkPublicBufferedEntityCommitPreparation retains the timed owner and adds untimed physical/visible node counts, acknowledged operation counts and natural pending-generation shape assertions. The private conflict-scan benchmark remains separately available. Samples ran with no other Go/model/container processes observed; OS background processes were sampled. Separate fixed300x CPU profiles are coarse and include inherited preparation/abort labels; they do not identify a universal bottleneck. The previously reported busy-host comparison is unqualified.