Context
test/playwright/unit/ai/services/memory-core/MemoryCoreRecorderService.spec.mjs:704 — independently downgrades the identity projection when its writer status is partial — flaked for the third time on 2026-08-12, each time recovering on retry while costing a full red/rerun cycle on an unrelated PR.
Evidence
- Third occurrence: run
31643257591 (PR #17040, a diff that does not touch Memory Core at all): first attempt failed expect(metrics.reembedRatio).toMatchObject({status: 'partial', reason: 'embedding-identity-writer-partial'}) at line 732, passed on in-run retry (1 flaky), and its worker collateral produced 1 error was not a part of any test + 1 did not run, turning a 13,000-pass run red.
- Two earlier same-day occurrences were absorbed as rerun-and-pass (recorded in Memory Core session context); the standing commitment was that a third occurrence earns this ticket.
The Problem
The assertion expects the reembedRatio projection to be partial at the observation moment; intermittently the projection reads ok, which suggests the fixture observes the projection before (or after) the writer-status downgrade settles — an ordering/timing dependency inside the fixture, not a production defect signal so far.
The Fix (owner to verify, not prescribe)
Make the arm deterministic: either drive the writer-status transition synchronously through an injected seam, or await the exact settlement signal the projection derives from before asserting. If investigation instead reveals a real race in the projection derivation, that becomes the finding and the fix moves to production code.
Acceptance Criteria
Related
PR #17040 (victim run) · epic-adjacent CI hygiene
Origin Session ID: 379c88ee-52c5-41ad-8973-8f28ebc8cbd6
Context
test/playwright/unit/ai/services/memory-core/MemoryCoreRecorderService.spec.mjs:704—independently downgrades the identity projection when its writer status is partial— flaked for the third time on 2026-08-12, each time recovering on retry while costing a full red/rerun cycle on an unrelated PR.Evidence
31643257591(PR #17040, a diff that does not touch Memory Core at all): first attempt failedexpect(metrics.reembedRatio).toMatchObject({status: 'partial', reason: 'embedding-identity-writer-partial'})at line 732, passed on in-run retry (1 flaky), and its worker collateral produced1 error was not a part of any test+1 did not run, turning a 13,000-pass run red.The Problem
The assertion expects the
reembedRatioprojection to bepartialat the observation moment; intermittently the projection readsok, which suggests the fixture observes the projection before (or after) the writer-status downgrade settles — an ordering/timing dependency inside the fixture, not a production defect signal so far.The Fix (owner to verify, not prescribe)
Make the arm deterministic: either drive the writer-status transition synchronously through an injected seam, or await the exact settlement signal the projection derives from before asserting. If investigation instead reveals a real race in the projection derivation, that becomes the finding and the fix moves to production code.
Acceptance Criteria
--repeat-each) after the fix.Related
PR #17040 (victim run) · epic-adjacent CI hygiene
Origin Session ID: 379c88ee-52c5-41ad-8973-8f28ebc8cbd6