Fix weight_data=True: NameError + arcsec/radian unit bug

[matthewholman]

This branch was developed with Claude Code, on top of an initial
implementation I had started.

Fixes two distinct bugs that have made orbitfit(weight_data=True)
non-functional since the feature was added. Either bug alone is
sufficient to break the path; both must be fixed for it to do
anything sensible.

Bug 1: NameError: name 'astcat_column_present' is not defined

_orbitfit in src/layup/orbitfit.py defined the flag as

g_column_present = "astCat" in column_names

but then referenced astcat_column_present two lines later inside
the debias and weighting branches. As soon as weight_data=True
(or debias) was passed, this raised NameError before any fitting
could happen.

Fix: rename the definition to match the usage sites
(astcat_column_present). One-line change.

Bug 2: arcsecond vs radian unit mismatch

data_weight_Veres2017 returns the astrometric uncertainty in
arcseconds (per its docstring). The call site assigned the
return value directly to Observation.ra_unc and dec_unc, which
are stored in radians. Each observation's weight was therefore
~(180·3600/π)² ≈ 4×10¹⁰× too loose -- inflating the fit covariance
to absurd values (sigma_pos ~25 million km on a 3 AU object) and
driving chi-square/dof to zero (residuals dwarfed by the assumed
noise floor).

Fix: convert arcsec → radian at the assignment, with a comment
explaining why.

Effect

Layup's reported position uncertainty for three real MPC objects
(mainbelt asteroids at ~2-3 AU, from the existing test fixture):

Object	weight_data=False	weight_data=True (after fix)
119839	269 km	123 km
742428	340 km	160 km
609631	306 km	80 km

With the fix:

• sigma_pos drops by ~half (closer to JPL's published per-element
  sigmas for these objects)
• chi-square/dof moves from artificially low (0.05-0.23, suggesting
  the default 1-arcsec uncertainty over-weights the data) to a more
  honest 0.27-0.53

Tests

tests/layup/test_weight_data.py (2 tests):

• test_orbitfit_weight_data_true_runs_without_crashing:
  guards against the NameError-class regression -- exercises the
  astcat_column_present branch by calling
  orbitfit(..., weight_data=True) on real MPC data with no
  astCat column.
• test_orbitfit_weight_data_true_gives_sensible_uncertainty:
  asserts sigma_pos < 10000 km for the 27-year mainbelt arc.
  Well above the actual ~123 km, but trivially below the
  25-million-km failure mode of the units bug.

Dependencies

Independent of all other open PRs -- branches off main, touches
only src/layup/orbitfit.py and adds one test file.

Review Checklist for Source Code Changes

• Does pip install still work?
• Have you written a unit test for any new functions? — tests/layup/test_weight_data.py
• Do all the units tests run successfully? — both new tests + existing suite
• Does Layup run successfully on a test set of input files/databases? — real-MPC fixture (4_random_mpc_ADES_provIDs_no_sats.csv)
• Have you used black on the files you have updated?