Graphify relationship transaction conflict

Current conclusion

The repeated dogfood ingestion failure is confirmed by retained workflow logs, but its root cause is not established. No production correction, data corruption finding, or completed reingest is claimed by this investigation checkpoint. The dogfood database and its volume have not been changed.

Release examined: 8f5835d7f2fbb9c1253a046ef4b9b64c2da06af1. This work is separate from CI PR #1736 and accepted-command draft #1691.

Observed failure

• Run 42020171895 and run 42108551854 fail importing the rolling release code graph into nornicdbcode.
• The latter logs 3,060,000 processed edges, then returns Neo.TransientError.Transaction.Outdated: SET r +=: conflict detected: edge nornicdbcode:merge-0624e1ecab8de2941dd9ea20a37bd4007b41153f2f283d37bb7facf33c46fb4e changed after transaction start.
• The reported relationship is IMPORTS_FROM, from pkg_search_rerank_mmr_extra_test to internal_testfixture_archive, scoped to repository orneryd/NornicDB.
• The operator reports the last completed ingest marker remains 35e3bddbf092f7cf9dd56ecfd15a32795cabc0d6. This has not been independently re-read from the dogfood database in this investigation.

The importer's progress counter also counts skipped unchanged edges. Therefore 3.06 million processed edges is not evidence that the failing request wrote that many relationships or that a storage-size threshold was crossed.

Relationship to existing work

NOR-288 and draft #1691, reference head 182893c0dd15db761ef604c7a6eefb285e4d7b64, own settlement after durable acceptance, native application conflicts, retained authority, semantic replay, and DROP.

The observed SET r += prefix identifies statement-time execution. Relevant current owners are BadgerTransaction.UpdateEdge, getCommittedEdgeForUpdateLocked, readEntityAtWithView, and getEdgeAtInTxn. A snapshot read reporting absent/not-visible followed by a live latest read is classified as a retryable conflict. That classification is intentional for a real post-BEGIN peer write. #1691 retains this classification; its accepted settlement loop is not a direct correction for the observed statement error.

This establishes different immediate paths, not proof that accepted-command history or another writer cannot contribute indirectly. No #1691 stack or codec change is copied into this work. NOR-213's broader writer lock-wait contract and NOR-371's replay-eligibility work are also not silently absorbed.

Isolated evidence

All probes use disposable native Badger storage through the public Cypher executor and the importer's exact statement:

UNWIND $rows AS row
MATCH (a:Code {id: row.src, repo: $repo}),
      (b:Code {id: row.tgt, repo: $repo})
MERGE (a)-[r:IMPORTS_FROM]->(b)
SET r += row.props

Probe	Result	What it establishes
Four rounds over 2,000 distinct relationships, 1,000 rows/request	Pass, 0.81 s test	Clean native batched creation and updates work; exact counts and property bounds checked.
One relationship repeated 1,000 times/request, three rounds	Pass, 0.13 s test	Duplicate pairs alone do not reproduce the failure; one final relationship and last-row values checked.
Surrounding 2,000-row batch from local release extraction, three rounds	Pass, native 0.98 s and WAL-backed 0.92 s	Local captured scalar payload works on clean native and WAL-backed stores; final count checked.
1,550 rounds over 2,000 relationships, 3.1 million total updates	Pass, 336.34 s test	Accumulated sequential updates alone do not reproduce it in this fixture. This is not 3.1 million distinct relationships or the dogfood history.
Six supported-history controls: memory, persistent retained/pruned history, edge deletion/recreation, endpoint deletion/recreation, reopen, head-only pruning with ID reuse	Parent independently passed 24-iteration matrix, 0.78 s test; source SHA-256 unchanged before/after	These supported serial histories do not reproduce the failure in the probe. They do not qualify dogfood history or concurrent writers.

The local release extraction contains 72,209 nodes and 3,734,421 edges. It places the watched relationship at processed edge 3,064,551, in the Code-to-Code IMPORTS_FROM bucket at position 2,928,715. This aligns with the reported failure region but does not reproduce the failure. The extraction differs from the CI population (71,730 nodes, 3,724,083 edges), including platform/header parsing differences. Stored hashes determine which rows the incremental importer writes; without that inventory the captured full bucket is not the exact failed HTTP request. No private or full dogfood replay is claimed.

Initial CGO-disabled Mac qualification did not build because of unavailable Metal symbols. Native CGO checks with -tags 'noui,nolocalllm' passed as listed; UI assets, local LLM integration, complete affected packages, make test, and make verify are not qualified by these selected probes. make verify-list only previews checks and is not a verification pass.

Probe source, captured rows, complete logs, and the independent DeepSeek source comparison are retained in the local task evidence directory. Passing diagnostic probes are not presented as a regression for a defect they have not reproduced.

Required next evidence

The active Mac Docker context and its default context currently list no running containers. The operator has been asked for the actual dogfood Docker host/context or a disposable native backup. A native copy must retain the MVCC head, version, prune-floor and sequence metadata; a Cypher-only export cannot establish the original history state.

On that isolated copy, capture the failing request's rows, BEGIN physical and logical read versions, current relationship head/floor, corresponding retained record, and concurrent writer activity. Distinguish a genuine intervening commit from absent/inconsistent historical metadata before choosing a correction. Preserve the real peer-write conflict and cancellation controls. Do not bypass the conflict, silently refresh a pinned snapshot, delete the relationship, or increase retry counts as a substitute for fixing the cause.

Reingest remains gated

Do not reset or reingest dogfood as an experiment. After the owning correction is reproduced and verified, review a single-writer incremental reingest against a frozen release commit, using the existing certificate/authentication and deletion safeguards. Only a completed run followed by a read of the repository's exact last_commit marker can establish completion. Neither a proposed run nor partial progress is a completed graph refresh.