reliability(patterns): atomic save of usage cache

[CryptoJones]

`socrates patterns review` on a CompanyOS with many projects can take
seconds to walk every .md file. A SIGINT during the final
`(patterns_dir / .usage-cache.json).write_text(...)` previously left a
truncated/invalid file. The next run's `_load_usage_cache` would
JSONDecodeError on the partial file and return None — silently
discarding ALL cached segments and forcing a full re-scan. (The current
load already swallows the JSONDecodeError into None; the bug is the
silent cache reset, not a crash.)

Fix: same atomic-write pattern as interview.py's save —
`.tmp` + os.replace. Cache is either fully old or fully new,
never half. Tempfile cleaned up on both success and failure paths.

Tests added (2):

• post-save: no .tmp leftover
• mid-save os.replace failure: pre-existing cache survives intact,
  no .tmp leftover

149/149 tests pass; ruff + mypy clean.

Self-review caveat: duplicates the atomic-write logic in #refactor/shared-atomic-write-and-decide-lock (which extracts it to _atomic.py). Merge order: refactor first → rebase this onto the shared helper. Functionally correct either way.