Benchmark curated vs full UMLS entity extraction

[AlexMikhalev]

Summary

Compare precision/recall of different entity extraction configurations to quantify accuracy/performance tradeoffs.

Details

Compare three configurations on the 18 evaluation cases:

1. 49-term curated thesaurus (current snomed_thesaurus.json)
2. 1.4M-pattern full UMLS automaton (current umls_automata.bin.zst)
3. Expanded curated set (500+ terms, from expanded thesaurus work)

Metrics to capture:

• Precision (correct extractions / total extractions)
• Recall (correct extractions / expected extractions)
• F1 score
• Extraction latency (ms per case)
• Memory usage per configuration

Acceptance Criteria

• Benchmark harness with reproducible results
• Comparison table across all three configurations
• Latency and memory measurements
• Recommendations for production configuration
• Results documented in repo

Priority: P3