Cell `i:branch_delete×change:distinct-branch` (crates/omnigraph-server/tests/server.rs ~4197) intermittently fails in CI's Test Workspace job: ``` assertion `left == right` failed: [i:branch_delete×change:distinct-branch] change left: 4xx (the /change response) right: 200 ``` The cell asserts that a `/change` on `main` succeeds while a *different* branch is concurrently deleted ("both should succeed"). Under contention the change's publish appears to hit a manifest CAS conflict caused by the unrelated branch delete and surfaces a non-200 instead of retrying. **Occurrences** (identical signature, both on PRs not touching the engine/server concurrency paths): - 2026-06-10 — PR #176 branch (passed on plain re-run) - 2026-06-11 — PR #186 branch (storage-primitives PR; engine concurrency untouched) **Two candidate resolutions** (one of them is right, the test is currently asserting whichever we *want* the contract to be): 1. The `/change` path should absorb/rebase manifest CAS conflicts caused by unrelated-branch deletes (the non-strict write queue already rebases other contention) — then the test is correct and this is an engine bug. 2. Concurrent unrelated-branch deletes legitimately surface a retryable conflict — then the cell's "both succeed" expectation should accept the documented conflict semantics. Per invariants.md, observable behavior is contract (Hyrum); whichever way this lands, the matrix cell should pin it deterministically (failpoint-sequenced rather than timing-raced) so the answer stops being probabilistic.