prompt: retire the three shipped Clipper follow-ups to complete/

[Jammy2211]

All three follow-ups merged today, so they leave draft/. Left in place they would read as unstarted high-priority backlog on the dashboard people pick work from — the exact drift this session opened by finding phase 1 still sitting there after it had shipped.

prompt → record	PR	merge
clipper-usage-in-search-summary	#1478	bbceff6
save-json-numpy-scalar-typeerror	#1479	b6e89cd
crashed-run-poisons-resume	#1480	5c9244b

Two corrections the records carry

Both are cases where the prompt asserted something implementing it disproved.

The search.summary prompt claimed there was "no search-specific channel at all". False. search_summary_from_samples already read samples.samples_info and already emitted the NaN counters and their rates, key-guarded so other searches are unaffected. The real gap was two specific omissions — n_clipped_lane_steps never leaving search_internal, and n_constrained_lane_steps reaching samples_info but never being printed — not a missing mechanism. Recorded because it nearly caused a rebuild of something that already existed.

The poisoned-resume prompt claimed the rerun was "a zero-step no-op that reads as a clean result". That did not reproduce. What reproduces is a hard JSONDecodeError on every rerun of the same search name, naming no file and offering no remedy. The no-op variant presumably needs a surviving search_internal whose restored total_steps short-circuits the loop — the crash path deletes it first. Same root cause, fix covers both, but only the crash is evidenced and the record says so.

Other things kept because they were learned, not planned

• A second unguarded JSON writer the float32 prompt never named — Samples.info_to_json, and the more dangerous of the two, since samples_info gains a counter every time a search does. #1478 added two to that very dict the same day.
• The resume-guard leg, found by auditing sibling code rather than from the prompt: the same narrow except-tuple meant a corrupt search_internal raised instead of falling into the fresh-start branch directly beneath it.
• A test trap — an end-to-end resume test passes while asserting nothing if the second run is short-circuited as already-complete, and the .completed marker is not found by rglob under the test config.
• The toy 3-parameter Gaussian that reproduces the autolens_profiling#128 mechanism in seconds (378 value-NaN lane-steps unclipped, 0 clipped, 414 clips at rate 0.958) — a far cheaper regression fixture than the imaging/mge cell.
• LBFGS does not poison. It simply refits and overwrites; the hazard is specific to searches that read prior output while resuming.

Bookkeeping

active.md updated to shipped state with both corrections. Four dangling references repointed at the new records. lifecycle.py check and index --check both OK; dashboard.md regenerated (141 → 138 drafts).

🤖 Generated with Claude Code

https://claude.ai/code/session_01FzF2XmKQaqZRWZfZMxvTNR

---

Generated by Claude Code