[gmloose]

Replaces the C++ extension _cbdsm (Boost.Python, MINPACK, PORT3) with a pure-Python module, bdsf/mgfunction.py. It provides the part of the MGFunction, Gtype and lmder_fit interface that gausfit.py uses. The unused fitters dn2g_fit and dnsg_fit, and the unused MGFunction methods for removing, getting and setting single Gaussians, were dropped. In gausfit.py only the imports change; mgfunction is imported at module level, so that the forked workers share it.

How it works

• Fitting: lmder_fit uses MINPACK's LMDER through scipy.optimize.leastsq, with the same settings and analytical Jacobian as the C++ code.
• Exact arithmetic: some island fits are ill-conditioned: a last-bit rounding difference can change a fitted Gaussian by tens of percent. A first NumPy version (one matrix multiplication for all Gaussians, float64 exp) made single fits about 2× cheaper than the final version, but it changed 2 of the 50 Gaussians of the test image (total flux by +67% and +10%), and test/tbdsf_process_image.py failed. The Gaussians and their derivatives are therefore computed with exactly the same floating-point operations as the C++ code, including the exponential in long double precision.
• Speed: the Gaussians are evaluated one by one, in place where possible (temporary arrays slow down fits of large islands). A single fit of a small island is still slower than in C++ (NumPy overhead per operation), but fit_island repeats most fits with identical input (74% of all fits for the test image, 84% for NGC6048). lmder_fit caches results per MGFunction, which gives exactly the same results.
• Errors: diverging fits give inf/nan as before, independent of global NumPy error settings.

Build

• Removes src/c++, src/minpack and src/port3.
• Removes Boost from CMake, cibuildwheel (before_all.sh, before_build.sh, environment), the CI workflow and the README.
• The Fortran (_pytesselate) and C (natgrid) extensions remain.

Results

On a 20-core i7-12700 with NumPy 2.4.6 and SciPy 1.18.1, against a build of master (7ddb350), 3 interleaved runs each:

	C++	Python
test/tbdsf_process_image.py	6.62–6.67 s	6.22–6.26 s
benchmark/run_pybdsf.py (NGC6048), total	63.8–64.1 s	54.2–54.8 s
… of which Op_gausfit	12.6–12.9 s	7.9–8.0 s
… of which Op_wavelet_atrous	40.1–40.2 s	35.5–35.8 s

• Outputs: all FITS outputs, catalogs and images, are bit-identical to master: 51 for the test image, 40 for NGC6048.
• Single fits (replaying all recorded fits, without the result cache): 904 of 904 fits of the test image and 33089 of 33097 of NGC6048 are bit-identical. The other 8 are 2 distinct fits (one repeated 7 times), which diverge in both versions: C++ ends with NaN parameters, Python with huge ones. Neither converges, so gausfit rejects both. Earlier, on an older master, such differences were traced to SciPy: since version 1.15, it uses a C translation of MINPACK, which in rare ill-conditioned fits rounds differently from the original Fortran; with the Fortran LMDER, built with f2py, those fits were bit-identical too.
• A single fit takes 1.18× as long as in C++ in total for NGC6048: 7.7× for islands with < 50 pixels, 0.90× for islands with ≥ 1000 pixels.

Memory (peak total PSS of the process tree, interleaved runs; measured on an older master, 4fc79a2, before the later simplifications of mgfunction, which don't change the peak memory of a fit):

	C++	Python
test/tbdsf_process_image.py (5 runs)	136–138 MiB	136–137 MiB
NGC6048 (6 runs)	mean 3759 MiB	mean 3759 MiB
… peak during Op_gausfit of the image	1404 MiB	1342 MiB

• The peak of NGC6048 is in the wavelet stage, in the main process, which doesn't use the fitting code; it varies by ±50 MiB between runs of either version.
• Importing mgfunction in each worker copied about 2.5 MiB of shared pages per worker (C++: 1.5 MiB for _cbdsm); a module-level import avoids that.
• A fit of the largest NGC6048 island temporarily needs a few MB more in Python, because SciPy copies the Jacobian; Python frees it after each island, while C++ keeps its caches until the worker ends.

Testing

• New test/tmgfunction.py, also added to CI and to the cibuildwheel test-command. It covers:

  • the residual and the Jacobian (against finite differences) for mixed Gaussian types
  • fitting, and fixed parameters
  • cached fits
  • find_peak
  • the Python interface and invalid input
  • degenerate fits: no exceptions or warnings, also with seterr(divide='raise')

• pip install . from a clean checkout builds without Boost; tmgfunction.py, tmulti_proc.py and tbdsf_process_image.py pass.

• benchmark/replace-cbdsm/ contains:

  • record_fits.py: records all fits of a run, with _cbdsm or mgfunction
  • replay_fits.py: compares the results, speed and memory of single fits of two implementations (a Python file, such as an older mgfunction.py, or a module such as bdsf._cbdsm)
  • run_comparison.sh: an end-to-end comparison that checks every FITS output exactly

  Details are in its DESCRIPTION.md.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-08T14:02:59.548015Z	902bc59	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 06e9d7e946

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".