fix(agent-cli): Windows temp-cwd cleanup must not fail a successful batch

[ppcvote]

Closes #315.

Root cause — one correction to the issue's hypothesis

It isn't a cross-worker collision on a shared temp file: every invocation already gets a unique directory (tempfile.TemporaryDirectory(prefix="skillspector_cli_") → mkdtemp under the hood). The race is delete-at-__exit__ vs the CLI's process-tree teardown. On Windows, a directory cannot be removed while any live process has it as its working directory or holds any handle inside it — I verified both hold modes raise exactly WinError 32 in controlled experiments. The claude binary on Windows is a .cmd shim that spawns node children; when any of that tree outlives the direct child (or a killed-but-unreaped child lingers on the timeout path), rmtree at __exit__ throws after the model has already answered, and llm_analyzer_base records the batch as llm_batch_failed. Concurrency widens the teardown window, which is why the issue's batch runs fail while one-file-at-a-time runs pass, and why the failing set varies between runs.

Fix

mkdtemp + explicit best-effort cleanup in a finally: _cleanup_temp_dir retries briefly (10 × 0.2 s — enough for the common case where the holder exits within a beat), then falls back to ignore_errors and leaks the directory with a warning. A cleanup failure never outranks a successful response. Detection/scan behavior is untouched; concurrency is preserved (no serialization).

Verification on Windows 10

Mechanism-faithful repro: a fake claude.cmd that answers normally on stdout but detaches a grandchild whose CWD is the temp dir (via Start-Process, so no pipe-handle inheritance — matching how the real CLI's children behave), 8 concurrent batches through run_agent_cli:

build	result
main	8/8 batches fail, PermissionError: [WinError 32] ... 'C:\...\skillspector_cli_<rand>' — the exact signature from #315
this branch	8/8 batches succeed; dirs reclaimed by retry when the holder exits quickly, else leaked with one warning

Controlled experiments (independent of the shim) confirm each mechanism: another process's CWD blocks rmtree (WinError 32 naming the directory — matching the issue's error, which names the directory rather than a file); an open file handle inside blocks it too (naming the file).

Tests

5 new cases in tests/unit/test_agent_cli.py: happy-path removal, transient-failure retry, never-raises-when-stuck (deliberate leak), a Windows-only real open-handle hold, and a regression test asserting run_agent_cli returns the model's response even when rmtree keeps failing. 87/87 pass on Windows — which also exercises the suite on the platform this provider was breaking on (4 pre-existing, unrelated Windows failures elsewhere in the full suite reproduce identically on a clean main; detailed in #316).

🤖 Generated with Claude Code

[rng1995]

[Automated SkillSpector Review]

Approved. Explicit best-effort cleanup prevents a Windows temp-directory handle race from converting an already-successful CLI response into a failed batch. Cleanup still retries transient failures, then warns and leaks rather than raising; the directory remains unique per invocation and cleanup runs in finally. The agent-CLI suite passes (86 passed, 1 Windows-only skipped).