BSSUnfold

Added

• Numba JIT-compiled inner loops (_numba_jit.py) for iterative solvers:
  • @njit(cache=True) compiled functions with automatic disk caching
  • Graceful fallback to pure Python when numba is not installed
  • JIT functions: _doroshenko_inner, _kaczmarz_inner, _mlem_inner, _gravel_inner, _compute_log_steps_jit, _dose_weighted_mse_jit
• numba added as optional dependency (>=0.65.1)

Changed — Performance

• Doroshenko solver: 50x speedup (40.6 ms → 0.8 ms) — element-wise inner loop eliminates per-coordinate numpy overhead
• Kaczmarz solver: 14x speedup (1.4 ms → 0.1 ms) — JIT-compiled row update loop
• MLEM solver: 7x speedup (2.7 ms → 0.4 ms) — JIT-compiled multiplicative update
• GRAVEL solver: 3x speedup (~2 ms → 0.6 ms) — JIT-compiled weighted geometric mean update
• Monte Carlo uncertainty: pre-generates all noise vectors at once instead of per-sample dict creation
• Comparison metrics: _compute_log_steps and dose_weighted_error use JIT-compiled helpers when numba available

Fixed

• total_flux_ratio returned sum(reference)/sum(test) instead of sum(test)/sum(reference) per docstring

Improved

• Extracted _compute_log_steps DRY helper in comparison.py (was duplicated in 3 functions)
• Extracted _handle_extrapolation DRY helper in interpolation.py (was duplicated in 2 functions)
• 110 new tests in tests/test_improvements.py (validators, converters, matrix utils, Monte Carlo, dose calculation, interpolation, comparison metrics, EURADOS metrics, Detector integration)
• Test suite: 910 tests (was ~800)