Fit the RMSF over a sized window cut at the main lobe

[AlecThomson]

fit_rmsf fitted its Gaussian over only about sigma/2 either side of the peak, with the window sized from the analytic guess. That measures how sharply the top of the lobe curves, not its width at half maximum. On a lobe that is not Gaussian it read 2% narrow (flat in frequency) to 7% wide (boxcar), and the answer also moved with n_samples and with the guess.

Change

• fit_rmsf takes fitting_size (default 1.25): the window is that many analytic FWHMs wide, like WSClean's -beam-fitting-size, and stops one sample short of the lobe's first minimum on each side. However wide it is set, no sidelobe gets in.
• rmsf_fitting_size is exposed on FDFOptions, get_rmsf_nufft and run_rmsynth.
• rmsynth_3d and rmsynth_3d_from_fits gain do_fit_rmsf and rmsf_fitting_size. With fitting on, fwhm_rmsf_radm2 comes from one fit to the shared rmsf_arr (about 2 ms per cube) instead of 3.8 / (lambda^2 range). Off by default, so 3D output is unchanged unless asked.
• Multiscale's per-scale restore widths also go through fit_rmsf, so they pick up the new window. test_multiscale.py passes unchanged.

Accuracy against the exact half-max width

Error at n_samples 3 / 5 / 10 / 20 / 100:

RMSF	before	after
Gaussian	0 / 0 / 0 / 0 / 0%	0 / 0 / 0 / 0 / 0%
boxcar in lambda^2	+4.5 / +3.4 / +5.4 / +6.2 / +6.6%	-4.8 / -1.4 / -0.4 / +0.2 / +0.1%
flat in frequency, 800-1800 MHz	-2.3 / -2.2 / -2.2 / -2.1 / -2.1%	+1.9 / -0.6 / -1.0 / -0.6 / -1.1%
SB56289-like, 44 channels blanked	+2.1 / +1.5 / +2.6 / +3.1 / +3.4%	+2.1 / +0.6 / +0.6 / +0.7 / +0.6%
same, bottom 15 channels blanked	+2.2 / +1.9 / +2.5 / +2.8 / +2.9%	+0.3 / +0.7 / +0.9 / +0.8 / +1.0%

Tests

• New: exact on a Gaussian RMSF, within 1.5% of the boxcar half max at n_samples 10 and 100, a huge window gives the same answer as one that just reaches the first minimum, rmsf_fitting_size <= 0 raises, and the 3D fit matches fit_rmsf on rmsf_arr.
• test_options_defaults.py covers the new defaults.
• Locally, test_options_defaults, test_synthesis_utils, test_nufft, test_multiscale, test_rmsynth and test_tools_3d_dask pass (410 tests), and so do the rmclean_1d and rmsynth_1d notebooks. test_rmclean_graph_build_stays_linear_in_chunk_count failed once under four xdist workers on four cores, then passed serially with and without this change. It only times RM-CLEAN graph building, which this change doesn't touch.
• prek run --hook-stage manual --all-files is clean.

flint-crew/flint will expose do_fit_rmsf and rmsf_fitting_size once this is released.

🤖 Generated with Claude Code

https://claude.ai/code/session_01C5pn26DJht35yoKWbMJ1sW

---

Generated by Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.58%. Comparing base (1aa4680) to head (b43fa8e).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #98      +/-   ##
==========================================
+ Coverage   92.51%   92.58%   +0.06%     
==========================================
  Files          14       14              
  Lines        3207     3208       +1     
==========================================
+ Hits         2967     2970       +3     
+ Misses        240      238       -2     

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