Concurrent entity writer qualification

NOR-213 requires same-entity Cypher writers to wait and apply their mutations on the committed predecessor, rather than both staging stale images and failing one commit. The correctness regressions exercise node/relationship writes, rollback, cancellation, timeout and deadlock behavior, both parsers, and public Bolt/HTTP routes. Refreshed no-op images also participate in collection reads without staging a write; unrelated entities retain their existing snapshot rules.

Matched workload

Apple M2 Max, darwin/arm64, Go 1.27.3, GOMAXPROCS=2, GOGC=100, noui,nolocalllm. Baseline: release d844abc648a15bc185a42e1f36dc476cc94cd0f0. The same BenchmarkNOR213* source runs on both arms. Four invocations in ABBA order, each with three 300 ms samples, give six samples per arm/cell. Figures below are medians, not statistical significance claims. The candidate is the reviewed production source in this PR, recorded with per-file source and binary hashes in the qualification receipt. Task-owned tests and compilation were stopped before measurement; this is not a claim of complete machine isolation.

The contended fixture has 64 indexed nodes and 16 concurrent autocommit sessions. They increment one shared node or their own node. After every sample, the exact sum of stored counters must equal successful statements (b.N - conflicts). The uncontended fixture repeatedly updates one node from a single session.

Both storage modes use in-memory Badger. “Sync” means synchronous transaction publication, not disk/fsync durability; write-behind uses the in-memory buffer. These numbers qualify writer/ledger work, not disk WAL throughput or production ingest latency. Separate durable public-protocol tests cover those routes.

Workload	ns/attempt, base → candidate	B/attempt	allocations/attempt	conflicts/attempt
Uncontended update	73,857 → 75,192.5	44,846 → 45,437.5	419 → 422	none
Sync, one node	39,638.5 → 37,106	48,509 → 30,243	421 → 386	0.486 → 0
Sync, own nodes	40,522.5 → 42,455.5	45,110.5 → 46,712	428 → 449	0 → 0
Write-behind, one node	34,431 → 36,232.5	31,531.5 → 29,894	326 → 312	0.424 → 0
Write-behind, own nodes	45,729.5 → 36,687	31,004.5 → 29,337	333 → 312	0 → 0

Rejected statements count as attempts in the old implementation. Adjusting for successful writes, the shared-node baseline is approximately 77.1 µs/ACK in sync mode and 59.8 µs/ACK with write-behind, versus 37.1 and 36.2 µs respectively in the candidate. Thus the higher write-behind time per attempt does not mean lower successful-write throughput. Every candidate sample has zero conflicts and its counter oracle passes.

The required lock has a measurable cost: uncontended updates add about 592 B and three allocations; disjoint sync writers add about 1.6 KiB and 21 allocations (3.6% more bytes, 4.9% more allocations). This is not an allocation-neutral claim. Shared-node sync p99 also rises from 6.7 ms to 14.4 ms, while median statement latency falls from 75 to 68 µs; write-behind p99 is 13.5 → 18.4 ms with a 59 → 54 µs median. The PR does not claim FIFO fairness or tail-latency parity with fast failures.

Reproduction

GOMAXPROCS=2 GOGC=100 go test -tags noui,nolocalllm ./pkg/cypher \
  -run '^$' -bench BenchmarkNOR213 -benchmem -benchtime=300ms -count=3
GOMAXPROCS=2 GOGC=100 go test -tags noui,nolocalllm ./pkg/cypher \
  -run '^$' -bench 'BenchmarkNOR213ContendedNodeUpdate/sync/oneNode$' \
  -benchtime=2s -cpuprofile=writer.cpu -memprofile=writer.mem

The separate CPU profile contains fixture/setup and Go benchmark calibration as well as measured execution. It is dominated by runtime scheduling/parking and memory work; it does not establish that every ledger bottleneck is eliminated. Raw repeated samples, profiles, binary hashes and per-production-file source hashes are retained with the PR qualification evidence.