fix: sweep.py wipes search state by default to defeat .completed resume

[Jammy2211]

Summary

Run 322549 came back 3.5× faster than 322548 on the same A100 cell. Looked like cache magic. It wasn't — run 2 didn't sample at all. It loaded the cached samples.csv + Nautilus pickle from run 1 and reported the same total_samples=65500 with a meaningless time_per_eval_ms=2.82.

PyAutoFit's resume gate is the .completed sentinel file (abstract_search.py:520):

if not self.paths.is_complete:
    result = self.start_resume_fit(...)
else:
    result = self.result_via_completed_fit(...)

force_pickle_overwrite=True only re-writes output pickles on the resume path; it doesn't bypass the gate. The previous _samplers.py comment claiming otherwise was wrong, and PR #29's README inherited the same mistake.

For production (SLaM chained phases) the resume default is correct. For profiling it produces phantom speedups whenever a prior attempt completed sampling — including the post-fit-latent-crash case that PR #30 just fixed, where run 1's sampling completed before the crash and left .completed behind.

Changes

• sweep.py — --keep-completed flag (default off). When off, sweep removes output/searches/<sampler>/<ds>/<model>/<instrument>/<config>/ before each cell run, wiping .completed + Nautilus pickle + cached samples.csv. A one-line [clear-completed] removed ... log per cell so wipes are auditable.
• _samplers.py — fix the force_pickle_overwrite docstring; clarify it controls output-file re-writes, not the resume gate.
• README.md — rewrite the corresponding paragraph; document the sweep-level wipe as the actual mechanism.

Honest run-1 number

From job 322548's log timestamps:

• 10:55:35 script start
• 10:56:01 Visualization warm-up complete → JIT + setup = 26s
• 10:56:01 → 11:03:13 Nautilus sampling, 65,500 evals → 432 s → 6.6 ms/eval (real)
• 11:03:13 → 11:07:12 two final perform_updates + latent crash → ~4 min

The 2.82 ms in run 2's JSON is (load + viz wall) / cached-sample-count — not a per-eval cost.

Test plan

• sweep.py --help shows --keep-completed.
• Dry-run prints [clear-completed] (dry-run) would remove ... when the dir exists.
• Real wipe via _wipe_search_state(...) removes .completed + nested files.
• No-op when the output dir doesn't exist.
• Re-submit A100 with the wipe; confirm Nautilus iteration lines appear and per-eval cost lands around 6.6 ms.

🤖 Generated with Claude Code