Explicit transaction indexed-read qualification

Baseline: release/v1.5.0 c4ac624b88a3814c4d12d4014ddfe6cb3a95c0a5. Candidate: NOR-160 transaction population visibility plus pinned point-read correction. Apple M2 Max, darwin/arm64, Go 1.27.3, GOMAXPROCS=4, GOGC=100.

Workload and method

BenchmarkIndexedTransactionRead generates 512, 12,000 or 24,000 target-document :PDGraphNode:PDEntity nodes and 256 other-document nodes. It creates a unique id constraint and a document_id property index through public Cypher. MATCH (m:PDGraphNode:PDEntity {document_id:$d}) returns either count(m) or all nodes. The result cache is disabled in both arms. BEGIN/ROLLBACK and setup are outside timing. Every node result is consumed: count, document, logical id/value and an allocation-free uniqueness bitset establish the complete expected population.

Separate baseline/candidate binaries use the same benchmark source. Runs are sequential ABBA, three samples per invocation, 200 ms per sample, giving six samples per arm/cell. Reported values are medians; native identities are newly generated for each fixture, so physical layout is not byte-identical. The runs did not overlap this task's local Go tests or builds, and no local-model worker was used. CPU/allocation profiles were collected separately for the 12K explicit count read; profiles also include untimed fixture/calibration work and are not query-only CPU totals.

Explicit transaction results

Nodes	Result	Baseline ms/op	Candidate ms/op	Baseline MiB/op	Candidate MiB/op	Baseline allocs/op	Candidate allocs/op	Time / bytes / allocations
512	count	2.075	1.627	1.684	1.459	23,345	17,433	-21.6% / -13.3% / -25.3%
512	nodes	2.249	1.810	1.903	1.679	25,443	19,524	-19.5% / -11.8% / -23.3%
12,000	count	65.377	52.614	43.171	37.707	665,056	520,782	-19.5% / -12.7% / -21.7%
12,000	nodes	69.212	56.941	48.455	42.995	713,481	569,332	-17.7% / -11.3% / -20.2%
24,000	count	139.782	112.342	92.586	81.701	1,559,980	1,272,398	-19.6% / -11.8% / -18.4%
24,000	nodes	153.662	128.170	103.272	92.338	1,656,192	1,367,804	-16.6% / -10.6% / -17.4%

Autocommit controls ranged from -1.2% to +3.7% in median time; bytes and allocations differed by at most 0.1%. In a separate longer isolated 24K node-result control (ten samples per arm, 500 ms/sample), time changed from 65.459 to 66.314 ms (+1.3%), with 46,990,703 versus 46,999,472 B/op (+0.02%). These controls do not establish that every autocommit workload is unchanged.

Owning change and correctness limits

An ordinary native transaction already holds its physical snapshot and its namespace's registered begin-time version. When those versions still agree, its node point reads now reuse readNodeItemInTxn directly against that pinned snapshot instead of reading/decoding the logical head and registering another history reader for every candidate. The body decoder, version-qualified cache, embedding hydration and read-source provenance remain shared with other reads. The sampled generic history-read stack is absent from the candidate profile.

Rewound/zero versions, late namespace pins, absent physical snapshots and write-behind keep the existing logical/read-committed owners. Before/after owner-equivalence tests cover update/delete/recreate/create after BEGIN, missing nodes, sidecars/vectors, visible/stored reads, own writes/deletes, suppression and cancellation. The integrated visibility correction retains physical label/equality populations and applies live visibility at delivery. No storage payload, Badger key, WAL format or Go public signature changes.

This is a measured reduction, not autocommit parity. At 12K, the candidate explicit count read is still about 52.6 ms against 10.8 ms in autocommit. Body reads/copying and embedding metadata checks remain; this PR does not claim a snapshot-aware batch hydration implementation or resolution of the separate OPTIONAL MATCH adjacency and relationship-property index reports on NOR-160. The original private application dataset was not available; this qualification uses the documented generated shape. Local builds use noui,nolocalllm because the worktree lacks the native llama library; final native qualification is CI.