fix: preserve learned provider facts across concurrent saves

[ZeroPoint95]

Concurrent learned-provider saves could lose a fact at restart: an older save could copy only fact A, a newer save could persist A and B, then the older snapshot could replace that newer file with A. Atomic replacement kept valid JSON but did not preserve save ordering.

A separate writer mutex now covers the snapshot through atomic replacement within one process. Fact recording and readers keep using the short-lived memo mutex, and requests continue to schedule saves in the background. A regression drives actual concurrent learning and persistence, drains completed writes, and verifies every fact in a fresh store. The manual and performance documentation state the in-process guarantee and retain the separate-process cache limitation.

Controlled diagnostic overlays reproduce old-snapshot fact loss and verify the repair while learning and reads remain responsive. Diagnostic hooks are absent from the patch. Required local gates against merged dev base 401a1e032af93fecdc4bd299c84c6c22d5c9be4d passed, including a fresh provider suite in 68.143 seconds with every selected test passing on its first run. Focused race-detector checks and independent review passed; the canonical build passed on the publication candidate. This repair is separate from the test-only restart-fixture ownership repair.

Maintainer review and merge are required. Any staging attempt must use the final merged release record and its successful exact-SHA Full check. This PR performs no promotion; stable and production are outside scope.

Final candidate: a4804de6ce0d1d1e674e86c27f54f30c725eae57, including its assigned-number change entry for #1824. Independent final-entry review passed; production source matches the reviewed publication candidate.