v0.6.0

Robustness, correctness, and R/Python API-parity release. Closes the full review-and-fix sweep (issues #2–#16).

Correctness & robustness

• Rank-pool pooling now uses a fused embedding_bag (mode=sum), cutting peak training VRAM ~5.4 GB → ~0.7 GB at batch 16384 / max_species 256 — fixes the WDDM-spill driver watchdog TDR on species-rich targets (#6).
• CPU/CUDA hash parity: CPU hash_species aligned with the CUDA kernel's double-hash finalizer, so GPU-trained checkpoints score correctly on CPU (#7).
• Deterministic taxonomy vocab: genus/family IDs assigned in sorted order, removing CSV-row-order dependence that misaligned embeddings across rebuilds (#5).
• Deterministic MultiTaskLoss accumulator (scalar seed, device from ordered targets) (#16).
• Locale-free numeric parsing (parse_float_strict): rejects trailing garbage, comma-decimal-locale safe (#16).
• CSV header hygiene: strips UTF-8 BOM, throws on duplicate column names (#16).
• Reproducible per-epoch shuffle, exception-safe AMP scope, CSR offset validation, padding mask fix (#9, #10, #11, #15).
• None-unpack UB guard + GIL release in the nanobind model/predict paths (#16).

Inference & training surface

• Chunked predict bounds peak VRAM regardless of plot count; CPU-default inference (#2).
• Platform-aware CUDA allocator + auto-halve batch_size on OOM down to a floor (#3).
• First-class checkpoint evaluation: load_state, classification predictions, test/train fold accessors (#4).
• Checkpoint config/metadata loaders close the write-only asymmetry (#14).

R / Python API parity

• Thread categorical_ids through the R model methods (#12).
• Int64 ID accessors; pool_species_cap; categorical_vocab() on Dataset/Trainer/Predictor; from_csv_with_schema via resolve.dataset.csv(..., schemaSource=) (#16).

Packaging

• Repositioned as domain-general (compositional data, not ecology-only).
• Version synced to 0.6.0 across the Python, R, and C++ manifests.