feat(clean_slate): write-site dataset provenance, orphan reporting, gc phase, tests

[Jammy2211]

Summary

Final leg of the PyAutoMind dataset-bulk series: the wake_up cleanup now actually reclaims regenerated dataset bulk — safely, everywhere it runs, with tests. Closes #169.

The #167 rule (name+type literals in a simulator* file) missed every dataset written by a non-simulator script: the HowTo repos were a 100% miss and start_here.py-written dirs survived every sweep. The replacement judges by write site, never name mention:

• New bin/dataset_provenance.py — stdlib AST analyzer, order-sensitive: it tracks which dataset path each variable holds at each point and whether that variable reaches a write call (output_to_fits, json.dump, open(..., "w"), or a same-repo helper whose body writes the parameter — one level of interprocedural resolution for the simulators/util.py idiom). Verdicts: REGENERABLE (deleted), DOWNLOADED (network-binding files blanket-protect their datasets — real cached data, kept silently), ORPHAN (no writer — kept and reported for a human call). Deletion requires positive evidence; reads cannot launder into writes (derivations are followed only through path-shaped expressions — the sdp81 start_here.py binds the real dataset it reads three lines from the simulated one it writes, and the analyzer keeps them apart). aplt.fits_array is deliberately not a write: dropping one auxiliary product into a real dataset folder is not generating the dataset (this exact case would have condemned cosmos_web_ring, real JWST data).
• clean_slate.sh phase 1b delegates to the helper and hard-fails if it errors (no fallback — guessing provenance is what this replaces); prunes emptied dataset dirs; size warnings aggregate per dataset dir; new untracked-.ipynb_checkpoints sweep (__pycache__ deliberately kept — import cache); new per-repo git gc --auto phase (the HowTo repos had 42 MB of never-packed loose objects before this week's one-off gc).
• wake_up.md guardrail wording made honest: clean-slate deletes untracked regenerable datasets (recreated on demand); it is recoverable-by-regeneration, not "reversible".

Validation

• Tests 2 → 14 for clean_slate (phases 1/1b/2/3 previously had zero coverage): never-delete-tracked, simulator + start_here writers deleted, sdp81 read-vs-write doctrine case, download protection, orphan reported-not-deleted, per-dir size warning, empty-dir pruning, checkpoints swept / pycache kept, dry-run inertness, helper hard-fail aborts the sweep, unparseable-script tolerance. Full suite 178 passed.
• Real-workspace DRY_RUN=1 acceptance: newly catches simulated_lens/simulated_galaxy/csv_api_example/start_here_example/howtolens/howtogalaxy (~8 MB across repos today); reports dataset/imaging/tutorial as orphan; none of the never-touch set (smacs0723, a2744, weak/a2744_pyrrg, multi/rxj1131, sdp81, uv_wavelengths) classifies as regenerable.

Generated by the PyAutoLabs agent workflow.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AcPSpRG8YqseMwRdJEK28m