fix(contradiction): stop Path C orphaning a conflicted row (09/22 L-12) by arkash20 · Pull Request #1690 · caura-ai/caura

arkash20 · 2026-09-22T12:04:08Z

Row

oss-0814-l-12 — "Path C forward detection gates ALL supersession-edge writes on if not found — a canonical match after a flipped match leaves the older candidate conflicted but orphaned from the chain, the exact corruption shape Path A explicitly guards against."

Reproduced, and it is worse than the row states: both mixed-direction orderings orphan a row, not just flipped-then-canonical.

The bug

detect_contradictions_by_entities_async's forward loop wrapped both attribution branches in one if not found:, while the "conflicted" status write at the top of the loop stayed ungated. The first confirmed conflict in a run therefore set found and suppressed every later supersession edge — but the later candidates were still marked conflicted.

candidate order	edge that never lands	row left `conflicted` with nothing pointing at it
flipped → canonical	`new_memory → older_cand`	`older_cand`
canonical → flipped	`newer_cand → new_memory`	`new_memory`

Either way the row is neither active nor properly superseded: retrieval penalises it as conflicted, and no chain edge reaches it, so nothing can explain or undo it.

Actual pre-fix payload for the flipped-then-canonical case (from the new test):

{newer_cand: {status: active, supersedes_id: new_memory},
 new_memory: {status: conflicted},              # no outgoing edge
 older_cand: {status: conflicted}}              # orphan

How Path A guards it

Both Path A loops (_rdf_conflict_pass and the semantic loop in _detect) split the guard in two:

new_memory_is_outdated — the status-reversion guard. Stops a later canonical iteration writing new_memory back to "active".
supersedes_id — the chain-edge guard, gating only the canonical branch. The flipped branch is ungated, because many newer rows may legitimately supersede one older row.

The RDF loop's comment names the exact failure: "otherwise the older canonical candidate is left orphaned (outdated but unreachable via the chain)", and test_mixed_conflicts_complete_three_way_chain pins it. Path C had collapsed both guards into found.

The fix

Path C now uses the same split — new_memory_is_outdated unchanged, plus a separate chain_supersedes_id gating only the canonical branch. No new mechanism; the branch bodies are byte-identical, only their guards moved.

chain_supersedes_id is seeded from the row's current supersedes_id, as the semantic loop's is. Worth noting because the in-file comments claim a storage-side CAS that does not exist: memory_update_status has no supersedes_id IS NULL guard — its only CAS is the explicit expected_supersedes_id, which the forward paths never pass. Without the seed, a canonical write would silently re-point a row that already supersedes something and orphan its previous target.

reg-a73 interaction

The tracker flagged these as adjacent, and they are — in the same direction. Bulk ingest is the main real-world producer of multi-candidate mixed-direction Path C runs: dozens of rows about one subject land within seconds, so each row is judged against siblings both older and newer than itself. write.bulk_subject_batching reduces that to the last row per subject (every sibling then older, all-canonical), so with the flag on the orphan shape is rarer. It is off by default, and the flag does not remove the shape — a batch still lands against a pre-existing store. No conflict between the two changes; this one makes the un-batched path correct rather than depending on the flag.

Tests

tests/test_l12_path_c_canonical_edge_orphan.py — 5 tests asserting on the resulting chain (status + supersedes_id per row) read off the batch_update_status payload, not on source text. Includes an invariant helper: no row this run marks conflicted may be left without an inbound edge.

Revert-tested against the committed blob — with main's detector restored and the new tests kept, 3 of the 5 fail:

FAILED test_flipped_then_canonical_completes_the_three_way_chain
FAILED test_canonical_then_flipped_completes_the_three_way_chain
FAILED test_existing_edge_on_new_memory_is_not_re_pointed
3 failed, 2 passed

The other 2 pin behaviour the fix must preserve (one outgoing edge per row; the flipped branch's existing-supersedes_id guard).

pytest -k "contradiction or supersession or path_c or pathc": 367 passed, 1 skipped on this branch vs 362 passed, 1 skipped on main (+5 new). No failures on either side. mypy on the file: same 6 pre-existing errors as main, none new. ruff check / format --check clean at CI's 0.16.2.

Out of scope

A row whose single outgoing edge is already owned (by Path A, un-retracted) still leaves a second canonical candidate conflicted-and-orphaned. That is the chain's one-outgoing-edge design and Path A has the identical shape — a separate question from this row.

🤖 Generated with Claude Code

github-actions · 2026-09-22T12:04:20Z