feat: contradiction resolution ledger + undo across CLI/HTTP/MCP (#1172)

[ajianaz]

What

Fase 2 of #1172: contradiction resolutions become a first-class, auditable ledger.

• Uteke::contradiction_resolutions(namespace, limit) — the resolution ledger: superseded-but-not-restored memories with reason + timestamp. Membership is EDGE-driven (deprecated row JOIN memory_edges ON edge_type = 'superseded_by') — the same predicate undo_supersession resolves against, so listed ⇔ undoable.
• Uteke::undo_supersession(id) — now atomic: restore + edge-pair removal in ONE transaction (previously promote() committed first, so a failed edge delete left an ACTIVE memory carrying a live superseded_by pair). Vector-index re-add + recall-cache invalidation run as post-commit side effects, mirroring promote()'s contract.
• supersede() re-supersession fix — superseding an already-deprecated memory now refreshes deprecate_reason/deprecated_at, so the ledger always names the CURRENT winner (previously the stale first-winner reason contradicted the edge pair).
• Surfaces (parity across all three):
  • HTTP: GET /contradictions?namespace=&limit=, POST /contradictions/undo ({id} → 404 when nothing to undo)
  • CLI: uteke contradictions list [--namespace --limit --json], uteke contradictions undo <id|prefix>
  • MCP: uteke_contradictions, uteke_contradictions_undo
• Fix: the no-namespace ledger query declared LIMIT ?2 while binding a single parameter → runtime bind error on every call. Caught by the new MCP roundtrip test.
• Docs: docs/cli-reference.md (contradictions section), CHANGELOG [Unreleased], docs/api-reference.md regenerated via docgen.

Why

#1172 Fase 2. Supersede was a side effect with no audit surface: agents couldn't see what the pipeline decided or revert a wrong decision without hand-editing edges. The ledger + undo close that loop, and the edge-driven membership guarantees the ledger can never disagree with what undo actually operates on (cora review caught exactly that divergence — plus the non-atomic undo — across three review rounds; all three findings are fixed here with regression tests).

Testing

• cargo test --workspace — 554 passed, 0 failed (incl. 4 new: core resupersession-ledger regression, MCP list/undo roundtrip with short-id resolution + double-undo error, 2 server API tests via TestRequest: ledger-list/undo-restores + undo-404/malformed-400)
• cargo fmt --all -- --check clean; cargo clippy --workspace --all-targets zero warnings
• cora review --staged pass (3 iterations: fixed non-atomic undo, stale re-supersession reason, LIKE-based ledger membership — each verified by the new tests)
• cargo run -p docgen regenerates api-reference with both endpoints

[github-actions]

🔍 Cora AI Code Review

✅ No issues found. Code looks good!

---

Review powered by cora-code · BYOK · MIT