Feature-substrate Python boundary + reproducibility demo

[logannye]

Summary

Follow-up to the feature substrate (PR #28): prove the pull — train a model on rosalind features output and show the inputs (and the trained model) are bit-reproducible, hash-verifiable.

• python/rosalind.py — a dependency-light boundary (stdlib + numpy): features(index, alignments, ...) runs rosalind features and loads the per-locus table into numpy. Meets Python ML pipelines without a pyO3/pyarrow build.
• examples/reproducible_features_demo.py (numpy-only) — builds a toy index+BAM, extracts features twice, proves byte-identical TSV + matching BLAKE3 receipts, trains a pure-numpy logistic regression on a genuine learnable task (is-purine from pileup counts), and shows the trained weights + predictions are bit-identical across the two extractions.

Verified locally (output in the findings doc): 983,011 loci, receipts match: True, held-out accuracy 0.9990, weights bit-identical: True, predictions bit-identical: True. The novel claim, demonstrated end-to-end: byte-identical features → bit-reproducible ML training inputs.

Spec: docs/superpowers/specs/2026-06-02-feature-substrate-demo-design.md. Findings: docs/findings/2026-06-02-reproducible-features-demo.md.

Honest notes

• numpy-only by necessity: this environment has numpy but no pandas/scikit-learn/pyarrow/maturin, so a pyO3/pyarrow binding is not verifiable here. The demo + boundary are fully runnable + verified in-env; the headline is reproducibility (the is-purine task is deliberately simple but genuinely learnable).
• Deferred (next step): the zero-copy pyarrow in-process binding (replacing the dead-end python_bindings stub) + an Arrow/Parquet egress — built and verified where maturin + pyarrow exist. This demo proves the pull before that investment.
• No Rust changed; the Rust suite is unaffected.

Test plan

• Demo runs end-to-end on the toy genome; output captured in the findings doc
• python -c ast.parse on both modules; cargo fmt --check clean; cargo test green; 0 warnings

🤖 Generated with Claude Code