Public API type checks

Topiary checks its table and cache boundary incrementally. Install the development tools with pip install -r requirements-dev.txt, then run ./lint.sh. The supported checker is mypy 2.4.0 with pandas-stubs 3.0.5.260914, pinned together for reproducible pandas signatures.

The current checked scope is:

• TopiaryResult tables, metadata, long/wide views and result operations;
• native TSV/CSV readers and writers, including nullable Metadata fields;
• CachedPredictor loading, querying and saving;
• Isovar hypothesis and patient-background export readers;
• native per-bond evidence conversion with from_cleavage_results.

tests/typing/public_tables.py checks a consumer that loads a table, changes views, queries a cache and saves the result. It also includes invalid calls whose expected errors must remain errors. Function bodies in the declared modules are checked, including functions without complete annotations. Imported source modules supply their available type information but are not yet part of this diagnostic scope (follow_imports = "silent").

Prediction-table cells, custom metadata and dynamically supplied predictor plugins still use Any where the input schema is open. Pandas-compatible indexing can return a Series or another TopiaryResult; use .df when the consumer needs pandas' full indexing API. Specific MHCTools/MHCgnomes imports have local import-untyped allowances because those upstream packages do not publish type information. There is no global missing-import ignore and no blanket module error suppression.

This is a development gate, not a package-wide typing promise. Topiary does not yet publish py.typed. Expand the named checked scope and qualify installed-package consumers before adding that marker. Runtime validation and error messages remain necessary for notebook and CLI inputs.