store: publish the manifest before pruning stale experts - #34

Problem

store.save deletes expert files that are no longer in the new manifest before it publishes that manifest. If the process dies between the deletion and the atomic manifest replace, the old manifest still names the deleted files and the checkpoint cannot be loaded.

This can happen on a prune, a shape-changing save, a crash, a SIGKILL, or ENOSPC.

Fix

Count stale expert files before publishing, then delete them only after the new manifest is durable. The deletion is placed after the existing manifest publication block, so this change is deliberately independent of the publication mechanism.

Stale extra files are harmless in the meantime: the loader reads only the files named in the manifest, and the next successful save removes them.

Compatibility with #32

This composes with #32, which adds a retrying _replace for Windows reader locks. It touches neither _savez nor the manifest replace call, and I verified the combined tree in both merge orders: tests.test_soak and tests.test_windows_runtime pass together (13 tests).

Files

• minagi/store.py — defer stale-expert deletion until after manifest publication.
• tests/test_soak.py — new regression tests for interrupted manifest publication, ENOSPC, retry, and prune ordering.

The manual fault-injection soak runner stays in my fork; it is not needed for this fix.

Validation

• Applied to current upstream aa9d53f and to Fix Windows training: UTF-8 logs, direct plot spawns, safe checkpoint replace #32's branch; tests.test_soak passes 3/3 on both.
• Combined with Fix Windows training: UTF-8 logs, direct plot spawns, safe checkpoint replace #32 in either order: 13 tests pass, 1 Windows-only test skipped on Linux.
• No manifest schema or loader changes.