[martyn]

Summary

• Tiny PyPI distributions (.whl and .tar.gz sdist): Add MANIFEST.in pruning tests/, reports/, benchmarks/, experiments/, docs/, examples/, ops/, scripts/, etc., shrinking the PyPI sdist from 1.7 MB (7.14 MB uncompressed) to ~343 KiB (~373 KiB wheel). Enforce < 600 KB size gates and root allowlists on both artifacts in .github/workflows/tests.yml and tests/test_public_package_imports.py.
• Pre-subtraction NaN masking in masked_mixture_nll: Mask observed via torch.where(mask.bool(), observed, 0.) before subtracting components[None] so unobserved NaN entries cannot contaminate gradients when observed.requires_grad is True. Export joint_reconstruction_loss, masked_mixture_nll, NeuralComponentImputer, and masked_quantile_transport_loss from top-level particlegan.
• Lint runner resilience: Fall back to shutil.which('ruff') when sys.executable -m ruff is unavailable and raise a clear RuntimeError instead of JSONDecodeError when Ruff exits with empty stdout.
• Documentation: Surface the bispf_shape_gates=False ablation readout (reports/prior-ablation-2026-10-25/README.md, Arm A4: 13/49 vs. 19/49 PASS) and link docs/bispf-input-domains.md (input_domain='continuous' | 'categorical') in README.md.