[J-Endres]

Closes #201.

Replaces the internal BIfTI phantom config, NIfTI loader and NIfTI writer with the published bifti package (>= 0.3.0).

Changes

• Dependency: adds bifti>=0.3.0; removes ~480 lines of internal code.
• Python: requires-python raised from >=3.10 to >=3.12, since bifti requires 3.12 (and numpy ≥ 2.5.1).
• phantom/nifti_phantom.py: now only re-exports the bifti classes. NiftiPhantom, NiftiTissue and ResliceConfig remain as aliases of BiftiPhantom, BiftiTissue and ResliceTo, so existing user code and brainweb generation keep working. mr0 additionally exports BiftiPhantom, BiftiTissue, ResliceTo, Patient, PhantomUnits and PhantomSystem (the last two were documented but not previously exported).
• TissueDict.load: uses the bifti loader. Behavioural differences:
  • resampling to reslice_to is density-weighted averaging instead of trilinear interpolation with slab averaging, so resliced values can differ slightly;
  • the phantom's patient position is applied to the affine (no patient → unchanged);
  • func mappings are parsed by bifti's restricted grammar instead of eval; complex B1+/B1- maps keep their phase.
• TissueDict.save: now converts to bifti.NumpyTissues and calls bifti.NumpyPhantom.save(..., constant_threshold=1e-5). Differences to before:
  • NIfTIs are named <json name>_<property>.nii.gz (e.g. phantom_B1plus_0.nii.gz instead of phantom_B1+.nii.gz), still one file per property with tissues stacked;
  • data is stored as float64 / complex128 (was float32 / complex64), so files are larger;
  • the JSON has no reslice_to and tissues are written in sorted name order;
  • $schema points to the official schema URL.
• Docs (api-nifti.md, api-phantom.md, api.md) and CHANGELOG updated.
• Version: bumped to 1.3.0 (Cargo.toml, Cargo.lock, pyproject.toml, CHANGELOG).

Testing

Run locally without building the Rust extension (CPU torch, bifti 0.3.0):

• TissueDict.save() → TissueDict.load() round-trip of a two-tissue phantom reproduces all maps and the affine exactly; .build() works.
• A phantom with complex two-channel B1+ round-trips exactly (phase preserved), with warnings promoted to errors.
• Two phantoms saved to the same directory get separate files.
• A phantom JSON saved by the old code loads with the new code.
• Reslicing to a coarser grid with patient=HFS gives the expected shape and affine; tissue densities still sum to 1.

The simulation test workflow and notebooks were not run locally.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NHKPBcYamTCDbaAVtBYRMR

[github-actions]

Simulation test: main (before) vs. this PR (after)

Sequence	Status	max mag NRMSE	max phase RMSE	max runtime Δ
mr0_DREAM_STE_seq	✅	0.00e+00	0.00e+00	+4.7%
mr0_DWI_SE_EPI	✅	0.00e+00	0.00e+00	+0.9%
mr0_EPI_2D_seq	✅	0.00e+00	0.00e+00	+5.1%
mr0_FID_seq	✅	0.00e+00	0.00e+00	+5.9%
mr0_FLASH_2D_seq	✅	0.00e+00	0.00e+00	+3.0%
mr0_GRE_to_FLASH	⚠️	–	–	–
mr0_RARE_2D_seq	✅	0.00e+00	0.00e+00	+7.8%
mr0_SE_CPMG_seq	✅	0.00e+00	0.00e+00	+14.4%
mr0_STE_3pulses_5echoes_seq	✅	0.00e+00	0.00e+00	+3.6%
mr0_TSE_2D_multi_shot_seq	❌	0.00e+00	0.00e+00	+32.3%
mr0_bSSFP_2D_seq	✅	0.00e+00	0.00e+00	+6.3%
mr0_diffusion_prep_STEAM_2D_seq	✅	0.00e+00	0.00e+00	-8.8%

✅ mr0_DREAM_STE_seq: All tests passed.

✅ mr0_DWI_SE_EPI: All tests passed.

✅ mr0_EPI_2D_seq: All tests passed.

✅ mr0_FID_seq: All tests passed.

✅ mr0_FLASH_2D_seq: All tests passed.

⚠️ mr0_GRE_to_FLASH: No comparison possible: [Errno 2] No such file or directory: '/home/runner/work/MRzero-Core/MRzero-Core/tests/simulation_test/ref_files/mr0_GRE_to_FLASH.seq.npz'

✅ mr0_RARE_2D_seq: All tests passed.

✅ mr0_SE_CPMG_seq: All tests passed.

✅ mr0_STE_3pulses_5echoes_seq: All tests passed.

❌ mr0_TSE_2D_multi_shot_seq: Runtime deviation too high (32.26%) at acc=0.60000

✅ mr0_bSSFP_2D_seq: All tests passed.

✅ mr0_diffusion_prep_STEAM_2D_seq: All tests passed.

Workflow run — raw .npz data is attached as artifacts.