[ZeroPoint95]

The learned-quirks restart fixture could fail after verifying a valid memo file: it forced one synchronous save but left the request's scheduled writer alive, then discarded in-memory facts. That writer could snapshot the deliberately emptied map and replace the verified file before reload, causing “a memo on disk must be believed at startup.”

Both restart fixtures now drain their owned scheduled writes before simulating restart. The mandatory-map reset also uses the existing mutex. All JSON-content, startup lookup, fake-endpoint and subsequent request-count assertions remain unchanged. This is a test-only repair; production code, request behavior, deadlines and limits do not change.

Controlled diagnostic overlays paused the actual scheduled writer before snapshot: the original ordering failed the exact assertion, while drain-before-reset passed unchanged assertions under the race detector. Required exact-base local gate and fresh provider suite passed; focused restart checks under the race detector, make build and independent review passed. Diagnostic hooks are confined to investigation overlays and are absent from this patch.

The original Full-check failure remains preserved. A separate production save-ordering defect was independently reproduced and has a separate unpublished repair; this fixture patch does not claim to resolve it. Staging remains blocked pending its disposition and an exact-candidate Full check. Maintainer review and merge are required; this PR performs no promotion.

The actual-number internal change entry for #1823 is included at final head 430d3508e4e33086e34eaabf2639ebe564f5335d. Final exact-base local gate and independent entry review passed; final-head required CI remains the readiness check.