Deploy app from main@a3254bf (no paper/ binaries)

Add unified strain catalog (100K rows w/ provenance) + selective weak supervision for pH

Add MediaDive-derived features (medium pH, NaCl, n_media) — all 4 targets improve

Fix _derive_salt to pick optimum entries — salt MAE 2.52 → 2.17 (-13.7%)

Expand training corpus to 46K strains: species-name → NCBI genome + isolation features

UI prep: pre-train phenotype heads + pre-score uncultured media

v2 results: ESM-2 t6 (8M, 20-protein sample) loses to v1 hand-crafted features on all 4 phenotype targets

Tier 1.5: uncertainty quantification via quantile regression in recommend.py

v2 scaffolding: ESM-2 embedding extraction + GPU runner doc

Phase E #2: scripts/recommend.py — single-genome → ranked media + phenotype CLI

Phase E modeling: per-medium classifiers + recommender training script

Phase E scaffolding: MediaDive integration + strain↔medium links

Fix GTDB column names + accession resolution for v226 metadata schema

Phase C scaffolding: GTDB candidate selection + uncultured prediction

Final cleanup: sync OVERNIGHT_SUMMARY.md + fix size display for small files

Make OVERNIGHT_SUMMARY.md write atomic (avoid race with regen loop)

Fix predictions parquet type mix + plumb feature_cols through eval

Harden post-featurize chain: each phase runs even if previous fails

Eval report enhancements: TL;DR + per-strain predictions + per-family error

Add eval report generator + training table persistence + group-col override

Streaming fetch+featurize pipeline + 6× pyrodigal speedup + GCA version resolution

Rewrite BacDive client for v2 public API (no auth required)

Scaffold v0: BacDive + NCBI ingestion, genome feature extractor, XGBoost baseline