Fix NaN in spectral TV regularizer on flat maps (epsilon_tv=0)

[rohandahale]

What

reg_tv_spec / reggrad_tv_spec produced NaN on a flat spectral map when epsilon_tv == 0 (the default since #310). A flat coefficient map is exactly the multifrequency imaging start — add_const_mf initialises a constant spectral index, so every neighbour difference is zero and the naive sqrt(0) leaves a 0/0:

• in the value (the autodiff path under JAX), and
• in the analytic gradient (the NumPy path).

Downstream this killed the optimizer: an HL Tau mf_order=1 run gave J: nan and reconstructed a flat blob instead of the disk. With the fix it converges to chi2_vis ~ 1 and nxcorr 0.999 vs the exact v1.2.5 result.

Fix

Double-where guard in both functions (autodiff-safe; a flat pixel gets gradient 0, not 0/0). Minimal and self-contained — only multifreq_imager_utils.py is touched, no new shared helper.

Test

New TestSpectralTVFlatMapFinite in test_gradients.py: every tv_* spectral regularizer stays finite on a flat map at epsilon_tv=0, checked on both the NumPy analytic gradient and jax.grad of the value. The existing S5 FD-parity suite uses a structured map with epsilon_tv != 0, which is why it never caught this.

Origin

Regression from #310, which flipped the epsilon_tv default from 1e-12 to 0. Backend-touching. May warrant a release cherry-pick like the #240 stv_pol_grad fix.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 47.44%. Comparing base (20e604e) to head (e8bd8eb).

Additional details and impacted files

@@             Coverage Diff              @@
##           dev-backend     #311   +/-   ##
============================================
  Coverage        47.43%   47.44%           
============================================
  Files               55       55           
  Lines            26979    26984    +5     
  Branches          4599     4599           
============================================
+ Hits             12798    12803    +5     
  Misses           12689    12689           
  Partials          1492     1492           

☔ 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.

[achael]

looks good -- main still has default epsilon=1e-12, so I think we just need to cherry pick this into dev as well.