[lucidprogrammer]

Two integration tests run two writers recording an observation at the same time and expect one to commit and the other to be refused with a typed error: impact_cascade_probes::concurrent_record_observation_two_root_serializes_with_typed_error_on_refusal and impact_analysis_engine_capabilities::engine_serializes_concurrent_record_observation_with_typed_error_on_refusal. When both commits reach the same pending invalidation key, the losing writer is refused as that key's unique violation (CgError::UniqueViolation { entity: "invalidation" }), a typed refusal of an atomic commit, but each test accepted only CgError::Engine refusals, so they failed intermittently (seen on context-graph PR #9 and on this PR's first run).

What changes

• Both tests also accept the duplicate-key refusal from the losing writer. Their data checks are unchanged: exactly one writer's invalidations are committed and stored.

Verification

• Each test passed 30 of 30 runs locally after the change.
• GitHub Actions (Linux) on bd2186c: https://github.com/context-graph-ai/context-graph/actions/runs/42770483214 — green, every job passed.
• Reviewed by a second model family (SHIP).