fix(cache): sweep orphaned models tmp files on refresh

[pierreraby]

Fixes #130.

Background refresh is fire-and-forget (void this.refresh()), so a host exiting between the temp write and the rename leaves an orphaned commandcode-models.json.<pid>.tmp behind with no cleanup on the next startup.

This change:

• makes the temp name unique per attempt (pid + timestamp + random) so concurrent hosts never share one,
• sweeps temp files older than 1h on the next successful refresh (recent files are kept so a concurrent host's in-progress write survives).

Verified locally: typecheck clean, test-models 25/25 (incl. a new sweep test covering stale-removal + recent-preservation), test-runtime 11/11, test-pricing 8/8, prettier clean.

[github-actions]

Dependency Review

✅ No vulnerabilities or license issues or OpenSSF Scorecard issues found.

Scanned Files

None

[github-actions]

Memory benchmark

Compared base 99e2c0e with PR head 7b5b44b on the same GitHub-hosted darwin runner. Lower values are better.

Metric	Base	PR	PR − Base	Change
Stable RSS	120.1 MiB ± 0.3 MiB	119.9 MiB ± 0.9 MiB	-0.3 MiB ± 0.9 MiB	-0.2%
Physical footprint	66.7 MiB ± 0.5 MiB	67.0 MiB ± 0.4 MiB	+0.1 MiB ± 0.4 MiB	+0.1%
Physical peak	77.8 MiB ± 0.4 MiB	78.3 MiB ± 1.0 MiB	+0.3 MiB ± 0.5 MiB	+0.3%

Interpretation: No metric shows a clear Base-to-PR difference beyond its measured run-to-run variation.

Extension overhead above pi baseline

Metric	Base overhead	PR overhead	Difference
Stable RSS	+5.9 MiB ± 0.8 MiB	+4.8 MiB ± 1.0 MiB	-0.3 MiB ± 0.9 MiB
Physical footprint	+3.5 MiB ± 0.4 MiB	+3.7 MiB ± 0.7 MiB	+0.1 MiB ± 0.4 MiB
Physical peak	+4.1 MiB ± 0.8 MiB	+4.6 MiB ± 0.8 MiB	+0.3 MiB ± 0.5 MiB

Values are medians of 6 alternating, paired runs. The value after ± is the median absolute deviation (MAD). Each process was sampled 12 times after a 2500 ms warm-up.

Environment: pi 0.84.4, Bun 1.4.0, darwin arm64. RSS comes from ps; physical footprint and peak come from macOS footprint.

This is a comparative signal, not a pass/fail threshold. GitHub-hosted runner noise can affect absolute values.

[karaaslanz]

Reviewed exact head 7b5b44b. The stale-temp sweep is bounded to old cache temp files, preserves recent concurrent writes, and uses unique per-attempt temp names. Exact-head typecheck and focused model/cache regressions passed locally, and the required CI/security checks are green. Approving.