Fix critical/high findings from whole-codebase audit

[hd152]

Summary

Fixes the 13 critical/high-severity bugs from a whole-codebase audit. Each fix has a regression test that fails on the old code. The one exception is #12's native path: that needs the old wheel to demonstrate, so only its numpy mirror was checked against the old code.

#	Area	Bug → fix
1	auto_settings	Presets that don't list a setting were ignored when blending it, so one preset's value applied to every target (e.g. galaxy coarse chroma 0 → 50 on a 97% star-field blend). Unlisted presets now vote for the baseline value; booleans take the value with the largest total weight.
2	frame_processor	Pool workers only got array masters. dark_exptime never arrived, so mismatched-exposure darks were subtracted unscaled on the default parallel path.
3	cli	d = masters['dark'] overwrote the target-directory loop variable, so hierarchical sessions with darks lost their info.json (Bayer, WCS, GPS, name).
4	photometry_core, color_calibrate	A 3-axis WCS from the (3,H,W) cube broke Gaia projection, so --photometry never produced results.
5	color_calibrate	median(r / median(r)) is always 1, so --color-calibrate never changed the image. Now uses median(expected / measured).
6	stacking	With --auto at 15+ frames, median/linear_fit/ivw/wavelet silently became an unrejected patch-weighted mean.
7	registration	match_stars_affine subtracted the calculate_shift seed instead of adding it, so seeded affine failed for shifts over ~5 px.
8	pipeline	Phase-1/2 checkpoint resume skipped target inference and the auto-advisor. A shared _infer_target_and_advise now runs on every path.
9	psf_deconvolution	The blind-PSF "RL refinement" flattened the PSF (peak 0.068 → 0.011). Removed.
10	difference_imaging	ZOGY kernel flux powers were swapped, so S_corr was not unit-variance when flux_new ≠ flux_ref.
11	time handling	Origin DATE-OBS is local time with TIMEZONE stored separately. The new utils.obs_time_utc_iso applies it (dispersion zenith angle, photometry airmass, light-curve MJD), the stacked header now keeps TIMEZONE, and a below-horizon zenith angle returns None.
12	stacking + Rust fold_pixel	Online sigma-clip (--stream) froze pixels seeded by one sample. They now accept samples until they have 3. astro_native 0.36.0 → 0.37.0.
13	frame_processor	The GPU master cache was keyed on shape/dtype and only checked on a new shape, so the next target in the same process reused stale GPU masters. It is now keyed on host-array identity and checked per frame.

Behaviour changes to be aware of

• Performance #1 changes --auto output on most targets, including settings baked into saved _config.toml files. Worth re-running a known real session before release.
• Fix CI smoke run failures #7 makes seeded affine matches succeed more often, so some sessions that were silently translation-only will now get rotation correction.
• The native crate must be rebuilt (0.37.0) for Bump signpath/github-action-submit-signing-request from 1 to 3 #12's native path.

Test plan

• pytest -q -n auto: 1723 passed
• ruff check . clean; tools/lint_conventions.py 0 errors (same 15 pre-existing warnings)
• Each new test verified to fail against the pre-fix source (for Bump signpath/github-action-submit-signing-request from 1 to 3 #12, only the numpy path)
• Re-run a real session with --auto and compare against the 2.3.0 output
• CI native job builds 0.37.0 and passes parity tests

🤖 Generated with Claude Code