[TMHSDigital]

Fixes #27.

The problem. Every write of a cache key went through the same key.tmp. Two workers answering cases with the same text raced on it, and on Windows the loser raised PermissionError [WinError 32] out of Cache.put. That aborted the run and lost every call already paid for: 18 of 20 runs in the review's repro.

What changed

• Cache.put writes through a temporary file of its own (tempfile.NamedTemporaryFile in the same directory), then os.replace.
• A write that still fails (a virus scanner holding the file, a full disk) is counted in the cache stats as write_errors and dropped instead of raised. The answer is already paid for, so a missing entry costs one more call on a later run, not this run.
• The runner holds a lock per cache key across look-up, call and store. Cases that share a key are answered once per run, and the rest are served from the cache instead of each being billed. Different keys never wait on each other.

Checked

• New tests failed on main and pass now:
  • 16 identical cases at workers=16 make exactly one adapter call, with 15 cache hits and no exception. On main this test reproduced the WinError 32 crash. It passed 20 runs out of 20.
  • With os.replace raising PermissionError, the run still returns every record, counts 5 write_errors, and leaves no .tmp files behind.
• The review's repro reports 0 crashed runs out of 20.
• ruff, mypy --strict, and the full pytest suite pass.

🤖 Generated with Claude Code