Fix polchisq psi parametrization and gmst_to_utc inverse

[achael]

Summary

Two latent bugs in obsdata.py + obs_helpers.py, plus an explicit warning for the multi-day aliasing limit. The two xfail(strict=True) regression tests landed via #249/#250 are now passing and have had their decorators dropped.

Bug 1 — Obsdata.polchisq psi parametrization

obsdata.py:1188 packed psi = arcsin(V/I), but pol_imager_utils.make_v_image defines V = I*rho*sin(psi). The imager and the chisq disagreed on what psi meant, producing an O(1e-2) self-consistency residue when evaluated on the generating image. Fixed to psi = arcsin(V/(I*rho)).

Bug 2 — obs_helpers.gmst_to_utc inverse

obs_helpers.py:976 used a hardcoded constant solar/sidereal ratio (0.99726956...) for a linear inverse, while utc_to_gmst called astropy per-sample. The mismatched rate accumulated to several minutes over a 24h obs.

Replaced with astropy-derived local sidereal rate (sample at mjd and mjd+1), plus a [0, 24) wrap on the sidereal delta so UTC lands in the canonical solar day of mjd. Roundtrip is now ~1e-11 hr for any obs that fits in one sidereal day (~23.93 solar hours).

Companion warning — multi-day aliasing

The >23.93h aliasing limit is fundamental: GMST mod 24 is many-to-one in solar UTC, and once data is in GMST the diagnostic signal is gone. Added a UserWarning in Obsdata.switch_timetype on the UTC->GMST branch only — the only direction where the unaliased span is still visible in the input. The inverse GMST->UTC has no signal to detect this, so no warning there.

Test plan

• pytest tests/test_obsdata.py → 169 passed, 1 skipped
• pytest tests/ (full suite, on prior unrebased commit) → 730 passed, 91 skipped
• The two previously-xfail(strict=True) tests (test_switch_timetype_roundtrip_utc_gmst, test_polchisq_self_consistency) now pass; their decorators are removed.
• Three new tests covering the multi-day warning (test_switch_timetype_warns_on_multi_day_to_gmst, test_switch_timetype_no_warn_on_single_day, test_switch_timetype_no_warn_on_gmst_to_utc).
• Multi-day stress: verified utc_to_gmst handles negative / >24 / multi-day inputs; gmst_to_utc returns UTC in [0, 24) of mjd with <1e-9 hr roundtrip for any single-sidereal-day obs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 26.87%. Comparing base (e852f37) to head (7ba4c0f).

Additional details and impacted files

@@               Coverage Diff               @@
##           dev-backend     #251      +/-   ##
===============================================
+ Coverage        26.85%   26.87%   +0.02%     
===============================================
  Files               50       50              
  Lines            25321    25328       +7     
  Branches          4261     4262       +1     
===============================================
+ Hits              6800     6807       +7     
  Misses           17821    17821              
  Partials           700      700              

☔ View full report in Codecov by Sentry.
📢 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.

[rohandahale]

@achael Looks good!

One trivial CI fix: lint is red on tests/test_obsdata.py:11:1: I001 Import block is un-sorted. The new import warnings needs to go above import numpy as np. ruff check --fix tests/test_obsdata.py closes it.

Two minor suggestions:

1. psivec = np.arcsin(...) is exact mathematically (V/(I·rho) ∈ [-1, 1] by construction) but a defensive np.arcsin(np.clip(..., -1, 1)) guards against floating-point overshoot at the boundary.
2. At fully unpolarized pixels (rho = 0), the fixed expression hits 0/0 → NaN. The old code dodged it by computing arcsin(0) = 0.

psivec = np.where(rhovec > 0, np.arcsin(np.clip(imstokes.vvec / (ivec * rhovec), -1, 1)), 0.0).
Returns the old arcsin(0) = 0

[achael]

updated following your review suggestions @rohandahale , merging now