prompt: tracer-fits-existence-guard shipped (autolens_workspace#503)

[Jammy2211]

Mind state for the task that shipped as autolens_workspace#504 (merge b19750c). Issue autolens_workspace#503 is closed.

What changed

• draft/bug/workspaces/bug_fix_the_tracer_fits_existence_guard.md → complete/2026/08/tracer-fits-existence-guard.md — advanced through active/ at issue-filing, then recorded on merge via lifecycle.py record, with ## Original prompt folded in.
• active.md — task registered at start_dev, parked at ship sign-off, then released on completion. Net-zero against main.
• autonomy_log.md — one calibration row for the --auto run.
• complete/index.md, dashboard.md, dashboard.html — regenerated.

The task

scripts/imaging/modeling.py guarded on files/tracer.json and then, inside that branch, read image/tracer.fits. Under TEST_MODE the reduced search writes the former but not the latter, so the guard passed and the read raised FileNotFoundError — leaving the release smoke gate permanently red since cfd5334c (2026-04-14). Fixed by guarding each read on the file it actually reads, matching the shape scripts/guides/results/start_here.py already used. scripts/multi_galaxy/modeling.py carried the byte-identical bug and was fixed in the same PR.

Two things the completion record captures for whoever hits this next:

• The default smoke profile hides the bug. PYAUTO_TEST_MODE=2 bypasses the sampler and writes no tracer.json at all, so the guard is never entered. Only PYAUTO_TEST_MODE=1 reproduces it. Anyone re-checking under the plain default will wrongly conclude it is already fixed.
• The guard shape is now settled. draft/maintenance/workspaces/read_through_issues.md asks for this teaching section to be added to the group/cluster/interferometer examples — guard-each-read-on-its-own-file is what should propagate, or the same bug ships to those scripts.

Autonomy

--auto launch at effective level supervised (prompt header safe, bug work-type cap supervised). It parked at ship sign-off as the contract requires; the human answered with /prm, which authorized PR-open, merge and close-out. Recorded as merged-unchanged.

Leg 4 of the ship gate — Heart — was NOT evaluated: pyauto-heart is unreachable from a web-github session. Stated on the PR, the issue and the calibration row rather than assumed green.

Note on the base

origin/main advanced by 7 commits mid-task. Merged in rather than rebased, per the repo's no-history-rewriting rule. dashboard.md and dashboard.html conflicted and complete/index.md auto-merged; all three were regenerated (pyauto-brain intake --apply dashboard, lifecycle.py index --apply) rather than hand-resolved, since a textual merge of these yields a stale record count and stale dashboard rows. lifecycle check and lifecycle index --check both pass (1124 records), and the net diff against main touches only this task's files.

---

Generated by Claude Code