[mcencini]

What changed

• apps.mobafit: fits by Levenberg-Marquardt one voxel at a time. Each voxel's damped system is solved directly, with its own damping and its own acceptance test, and converged voxels drop out through model.select. cg_maxiter, inner, alpha, alpha_min, redu and reference are removed. The CLI's --liniter is accepted and ignored.
• nlop.SignalModel: renamed to nlop.ModelOperator, matching blochsim.ModelOperator. It requires blochsim 0.0.12.
• Documentation: AGENTS.md, the apps API page, the nonlinear explanation page and two examples now describe the new method.

Why

Without an encoding operator, the linearized problem is one small system per voxel. A global conjugate gradient converges only as fast as its worst-conditioned voxel. It stops with those voxels still pulled towards the Tikhonov reference.

Measured on 1000 voxels of a multi-echo decay with T2, M0 and a noise-floor offset unknown (true offset 0.05):

Fit	Time	Median T2 error	Offset
Global conjugate gradient	8 s	6.0 %	0.086
Per-voxel Levenberg-Marquardt	1.2 s	2.4 %	0.0499

Validation

• Suite: BARTORCH_LIBRARY=… PYTHONPATH=src pytest tests/test_apps.py tests/test_cli.py tests/test_nlop_blochsim.py gave 238 passed, 2 skipped.
• Wider earlier run: an earlier run that also took in test_docs, test_docstrings and test_dependencies gave 290 passed, 3 skipped.
• New test: test_mobafit_finds_what_scipy_finds holds the fit to scipy.optimize.least_squares with rtol=1e-3.
• Lint: ./scripts/lint.sh is clean.
• Not available: no CUDA device.

Public API and documentation

• API: nlop.SignalModel is now nlop.ModelOperator, and apps.mobafit takes fewer keywords.
• Docs: updated as listed above.
• Release order: this follows pulserver/torchsim (blochsim 0.0.12) and precedes pulserver.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Qarafd1QwmJGNY8ZwdvwUY

---

Generated by Claude Code

[mcencini]

The wheel test step fails on No matching distribution found for blochsim>=0.0.12; PyPI has releases up to 0.0.11.

The pin belongs to this PR, because nlop.ModelOperator needs the API that pulserver/blochsim#41 trims. It resolves once blochsim 0.0.12 is released, and I'll re-run CI then. Until that release, any other check that installs the package will fail the same way.

---

Generated by Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.14815% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/bartorch/apps/_mobafit.py	97.93%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!