TiltCorrection and TiltSolution improvements

[hpparvi]

This PR

• improves the TiltSolution.resample method by propagating uncertainties (if available), propagating the image mask, and copying the NDData metadata to the resampled frame,

• improves the stability of the TiltCorrection.find_arc_lines method, and

• adds a baseline subtraction option to line_matching.find_arc_lines (also related to arc line search stability).

AI and LLM disclaimer: same as in #319.

[codecov]

Codecov Report

❌ Patch coverage is 97.40260% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 92.65%. Comparing base (4e330bc) to head (41f85a5).

Files with missing lines	Patch %	Lines
specreduce/line_matching.py	94.28%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #320      +/-   ##
==========================================
+ Coverage   92.52%   92.65%   +0.12%     
==========================================
  Files          18       18              
  Lines        2341     2409      +68     
==========================================
+ Hits         2166     2232      +66     
- Misses        175      177       +2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[tepickering]

Can fill this placeholder with this PR's number.

[tepickering]

Looks good overall. The changes that should be made before merging are to fix the NaN poisoning, pass the mask to _estimate_baseline, and decide where the baseline subtraction default is set. I also noted the inconsistency between this and #319 w.r.t. empty masks being False vs None. Either is fine, but consistency is good. Better per-row normalization and handling non-square images with disp_axis=0 are probably worthy of their own issues.

[tepickering]

Defaulting to True here changes more than TiltCorrection: WavelengthCalibration1D.find_lines calls this function without passing the flag (wavecal1d.py:322), so wavelength-calibration line finding now silently subtracts a baseline too. The changelog only says baseline subtraction is "on by default in TiltCorrection.find_arc_lines".

Two options, either is fine with me:

1. Default to False here and pass subtract_baseline=True from TiltCorrection.find_arc_lines. This keeps the existing wavecal1d behaviour unchanged.
2. Keep True, but add a changelog note about the change to WavelengthCalibration1D.find_lines, and expose subtract_baseline/baseline_window there so users can switch it off.

[tepickering]

The per-row normalization isn't NaN-safe, so one bad pixel poisons its whole output row. A quick check with a 20×30 frame and a single NaN pixel:

• mask_treatment="apply": all 30 bins in that row are NaN in both flux and uncertainty, but only 2 bins are masked.
• mask_treatment="nan_fill": the whole row is NaN and nothing is masked.
• An all-zero row (e.g. after zero_fill, or padding) gives 0/0, so that row is NaN as well.

This was already the case before the PR, but the new docstring promise that the mask marks every bin that received a contribution from a masked pixel makes it misleading now. Computing n from good pixels only would fix most of it, something like:

good = ~mask & np.isfinite(flux)
fgood = np.where(good, flux, 0.0)
den = (dtdx * fgood).sum(1)
n = np.divide(fgood.sum(1), den, out=np.ones(ny), where=den != 0)

There's a smaller version of the same problem in the multi-pixel branch: flux[~m, imin:imax] * k multiplies NaN by the zero weights outside a row's range, giving nan * 0 = nan, so a neighbouring bin can turn NaN without being flagged (the mask only counts w > 0). Using fgood there as well, or np.where(k > 0, flux, 0), would avoid it.

[tepickering]

The baseline estimate ignores spectrum.mask. Masked but finite pixels (saturated lines, bad columns) still go into the sigma-clipped median. Could we pass the mask through? sigma_clipped_stats accepts one directly:

baseline = _estimate_baseline(
    spectrum.flux.value, window=baseline_window, mask=spectrum.mask
)

and in _estimate_baseline, use sigma_clipped_stats(flux, mask=mask, ...) for the global median and mask=None if mask is None else mask[lo:hi] for each chunk.

[tepickering]

Small consistency point: here the output always carries a mask array, all False when the input had none, whereas WavelengthSolution1D.resample() in #319 returns mask=None when the input has no mask. Either convention is fine, but it would be nice if the two resample methods matched. Here the natural place to check is getattr(flux, "mask", None), read before parse_image, the same way meta and has_uncertainty are.

[tepickering]

One caveat worth adding here: n is computed from the data itself (flux.sum(1) / (dtdx * flux).sum(1)), so it isn't deterministic in practice. On sky- or background-subtracted frames, where a row sums to roughly zero, n becomes unstable, and since the propagated variance is scaled by n**2, the uncertainties go with it. Maybe a sentence like:

Suggested change

	Notes
	-----
	Each output bin is a linear combination of detector pixels,
	``F = n * sum_j k_j f_j`` with ``k_j`` the fractional pixel overlap times the
	Jacobian of the transformation and ``n`` the per-row flux-conservation factor.
	The variance is propagated as ``Var = n**2 * sum_j k_j**2 var_j``, assuming
	independent pixel noise and treating ``n`` and the Jacobian as deterministic.
	"""
	The variance is propagated as ``Var = n**2 * sum_j k_j**2 var_j``, assuming
	independent pixel noise and treating ``n`` and the Jacobian as deterministic. Since
	``n`` is estimated from the flux of each row, it is unreliable for rows whose total
	flux is close to zero, such as in background-subtracted frames.

A more robust per-row normalization could be a follow-up issue rather than part of this PR.