Data: config-driven ClinVarDataLoader and reference-validated windows

[cropsgg]

Summary

Refactors ClinVarDataLoader to take AppSettings: config-driven ClinVar fetch, reference-validated variant windows, and YAML-driven synthetic data (pathogenic templates, benign synonymous/intronic, VUS).

Base branch: pr/01-config-reference-assets (stacked PR — retarget to main after #1 merges, or merge in order).

What changed

• ClinVar parsing — Simple cDNA SNV HGVS patterns; optional RefSeq-in-title gate; reference base must match FASTA before a row is kept.
• hgvs_linear_overrides — Explicit0-based indices for HGVS (e.g. intronic) where automatic linearization does not match the bundled reference layout.
• Safety — Avoids labeling wild-type windows as pathogenic when a variant cannot be applied.
• Init checks — Optional rejection of ambiguous bases in the reference; validates synthetic pathogenic positions (and optional require_ref_base) against FASTA.
• Tests — tests/test_data_loader.py uses load_settings() + temp cache dir; keeps leakage/stratification checks.

Why this PR exists

Training data quality is the ceiling for biological plausibility. This ties labels to actual alleles on the configured reference and makes ClinVar + synthetic generation explicit in YAML.

Dependencies

• Builds on PR Config layer: Pydantic settings, default YAML, reference assets #1 (Config layer: Pydantic settings, default YAML, reference assets #1): settings + default YAML + reference files.

How to verify

pip install -r requirements.txt pytest
pytest tests/test_data_loader.py -q