[jeremy-wayland]

Fixes from testing the published 0.4.0 crates and pypasteur wheel against the synthetic-cohort sandbox models.

Fixes

1. JitterSimulator.transform before fit silently did nothing. feature_std started at 0.0, so an unfitted simulator returned the data unchanged, which an evaluation would score as perfect jitter stability. It is now Option<f64> and transform errors when it is unset. In Python this raises ValueError. The behaviour dates from 0.3.0. The CLI always fits first, so CLI results are unchanged.
2. pasteur-model README example didn't compile against 0.4.0. It passed 1 where from_file now takes Option<usize>. That README is the crates.io page.
3. Smaller issues:
   • pasteur-cli --version now works.
   • pypasteur.__version__ is exposed. It is listed in __all__, because maturin's generated __init__.py re-exports with import *, which skips dunder names.
   • Positive-class-index errors from the library now name the parameter as well as the CLI flag.

Changelog entries are under Unreleased. The workspace version is unchanged.

Testing

• cargo fmt --all --check, cargo clippy --workspace --locked -- -D warnings and cargo test --workspace --locked --exclude pypasteur-bindings all pass, run in a glibc 2.41 container. New tests: jitter_transform_before_fit_is_an_error and version_flag_prints_crate_version.
• Built a wheel from this branch. The updated tests/smoke.py passes; it now also checks __version__ and that an unfitted transform raises. A 13-check script covering the README's claims also passes.
• The 0.4.0 release itself passed the 78-check sandbox oracle suite (CLI from crates.io). The library crates' predictions matched onnxruntime to about 1e-15.

🤖 Generated with Claude Code