Fix progress bars, RCSB/FPSubSim2 robustness, and add dataset/API features

[OlivierBeq]

Summary

This PR brings master up to date with a large batch of work on dev: correctness fixes across download/conversion, RCSB matching, and FPSubSim2 similarity search; crash-safety and retry hardening for network/file operations; several new convenience APIs; and broad progress-bar/notebook UX fixes. It also adds substantial new test coverage.

New features

• Add PapyrusDataset.download for per-dataset descriptor downloads
• Add create_fp_subsim_search to the object-oriented (oop) API, documented in the advanced querying notebook
• Add collect alias for aggregate
• Add leave_level option to control QSAR/PCM progress bar behavior

Fixes

• Fix legacy papyrus++ key breaking the conversion progress total
• Fix read_protein_descriptors for a bare ProDEC Descriptor
• Fix PDB ID case mismatch in SIFTS mapping
• Fix RCSB cache path mismatch between write and read; switch RCSB cache to gzip (polars can't read xz)
• Skip RCSB ligands with unparseable SMILES
• Fix missing exit after an unknown fingerprint parameter in the fpsubsim2 CLI
• Default FPSubSim2.create() fingerprint to Morgan instead of building all ~10 fingerprint types (was a ~5.6x slowdown trap)
• Fix InChIKey/connectivity Binary decoding in FPSubSim2 mapping
• Fix dangling node close and consolidated progress bar in sort_db_file
• Fix pdbmatch progress total for papyrus++ datasets
• Fix progress bar stuck/clamped at 100% during descriptor and xz/gz conversion
• Prevent stuck progress bar widgets after early exit; always close progress bars in modelling.py
• Fix tqdm ncols squashing/width issues in notebook progress bars (download conversion, structure loading, MolSupplier, similarity/substructure filters)
• Remove MOE descriptor support (no longer available upstream)

Robustness & performance

• Make FPSubSim2 creation/sorting, pdbmatch output, RCSB cache, and links/aliases cache + gz/xz conversion writes crash-safe
• Retry POST requests and chunk-stream download failures instead of leaving truncated files/state
• Bound desc_type='all' descriptor joins via ids and streaming; bound chunk memory and widen the chunk cell budget for wide descriptor-table conversion
• Vectorize FPSubSim2 ID-to-mapping lookup for large hit counts; read parquet FPSubSim2 shards independently per worker
• Report file conversion progress by row count; pass known row counts to lazy conversion progress bars
• Skip re-downloading/re-extracting already-downloaded requirements archives; validate desc_type before downloading descriptors
• Delete already-downloaded originals on lazy conversion when keep_original_files=False
• Parse int-schema columns as float64 during conversion
• Raise RCSB GraphQL chunk size to the API limit
• Download structures before building the FPSubSim2 database (fixes build ordering)
• Replace deprecated RDKit fingerprint functions with generators
• Standardize neural network input features
• Require the cupy[ctk] extra for GPU support

Docs & notebooks

• Document create_fp_subsim_search in the advanced querying notebook
• Update and refresh outputs for the simple examples and advanced querying notebooks

Tests

• Substantially expanded coverage: FPSubSim2 end-to-end for every fingerprint type, fingerprint .get()/construction for every type, download chunk-stream retry/cleanup, keep_original_files/keep_xz forwarding, legacy papyrus++ key fallback, modelling progress-bar closure, neural network feature standardization, and more
• Narrowed/scoped exception handling and warning suppression in several test suites to their actual failure modes rather than broad catches

🤖 Generated with Claude Code