[joepetjr]

What

photos sort now builds one star trail per night session, after the stacks and before the time-lapse (follow-up named in #300):

• Output. trails/<first-group-id>.tif is the 16-bit master and trails/<id>.jpg is the browser JPEG, via night_products.render_master_preview. Each trail appends Output("trails", id, "trails/<id>.jpg"), so the feat(photos): one card per sequence on the sort index page #299 card shows it between the stacks and the video.
• Engine. It renders through the existing astro_trails.render_trail (calibrated, unaligned, streaming per-pixel maximum), behind the preflight_trail RAM/disk gate. A new GroupOps.star_trails seam wraps it.
• One triage per session. classify_night runs at most once per session, and both the trail and the time-lapse use its result. Lens-cap darks are never used as lights. Darks calibrate only when build_validated_master_dark accepts them; that check is now the shared _validated_darks helper. A maximum keeps an unsubtracted hot pixel forever, so an unmatched set is dropped.
• Threshold. TRAILS_MIN_FRAMES = 24 usable lights, about 13 minutes at the measured 32 s interval. A shorter session stores trails-skip:v1, so --resume doesn't decode it again.
• Resume and failure. These match the time-lapse: the ledger key is group_key("trails", members) + stack:v1. A failure becomes a skipped.csv row, its partial files are removed, and the run goes on. --no-stacks skips trails.

Real-data smoke (default seam, no stubs; 2026-11-09 run, night-1, output to a scratch dir)

• night-1: 19 RAWs → sg._star_trails → night-1.tif (107 MB) + night-1.jpg (3.1 MB) in 124 s at nice 19, beside a running photos sort --reprocess.
• I looked at the JPEG: even dashed arcs across the whole sky over the treeline and ridge, as the astro_trails docstring expects for this camera's 15 s exposures on a 32 s interval. It is not black and shows no hot-pixel field.
• The two real sessions are 76 and 94 frames, so roughly 8 and 10 minutes each at this rate.

Test plan

• Red: pytest -n 0 -p no:cacheprovider --no-cov -q tests/test_photo_sort_groups.py → 26 failed with TypeError: GroupOps.__init__() got an unexpected keyword argument 'star_trails' (tests/test_photo_sort_groups.py:118). The shared fake passes the new seam, and the 4 new tests failed too (no sg._star_trails, no trails/).
• Green: tests/test_photo_sort_groups.py → 34 passed.
• Final verification: test_photo_sort*.py, test_astro_trails.py, test_astro_timelapse.py, test_docs_numbering.py, test_changelog.py → 129 passed. ruff check and ruff format --check clean. ci/secret_scan.py clean.
• Refactor: resolved. The dark-validation block moved out of _night_timelapse into _validated_darks, which the trail reuses. Its existing unmatched-darks test still passes.

Regression verified; test-first order not independently enforced.