v0.2.1: Dual-model provenance

What changed

Sommelier API v0.2.1 closes the provenance gap between the two models served by the product.

• The regression lens now has an executable mln601-a1-derived-v1 contract linking the submitted MLN601 Assessment 1 protocol to the production adaptation.
• The classification lens remains locked to the submission-exact mln601-a2-v8 contract.
• /health, /model/info, the Streamlit Model Card, training, and parity checks now expose and enforce both lineages independently.
• Prediction schemas and model artifacts are unchanged from v0.2.0.

Regression lineage

The production regressor preserves the A1 estimator and features but is retrained after removing 1,177 duplicate rows. The contract makes that distinction explicit instead of presenting the serving metrics as the submitted result.

Protocol	Rows	R2	MAE	RMSE
A1 submitted	6,497	0.5002	0.4364	0.6075
Serving adaptation	5,320	0.4146	0.5096	0.6634

The regression golden prediction remains 5.065.

Classification lineage

The production classifier remains the A2 v8 class-weighted Random Forest selected by the declared cross-validation gates and interpretability tie-break. Its held-out test ROC-AUC remains 0.8337.

Verification

• Full suite: 51 tests passed.
• Contract parity suite: 30 tests passed.
• Regression artifact SHA-256: 10a18b80562a46c9233fc4456ca3d65b537524dd6c2c52c25336a43d0e61e7bd
• Classification artifact SHA-256: 15fb68f2dd3bd4534e03a7df6c48e124a9b9b5813b3f227fa328e57dc37d2609
• Input schema SHA-256: ee546ca46453f935513e46ce908521b9dee3bc44c588219608a14736dfe5a161

Full provenance and deployment checks are documented in the repository README and docs/launch/deploy.md.