"""Parsing stage: raw OCR text -> structured record.

Run in ocr_env. Input: data/ocr/*.json  Output: data/records/<form_id>.json

Field strategies:
- dates, blood pressure, weight, id numbers: regex with tolerant patterns
  (OCR confuses 0/O, 1/l, 5/S - normalize before matching)
- diagnosis: rapidfuzz fuzzy match against an ICD-10 code list CSV
  (data/icd10.csv - code, description; add Indonesian aliases as needed)
- every field carries a confidence; below a threshold -> needs_review: true

Output record shape is FHIR-flavored on purpose (Patient / Encounter /
Condition sections) - the same shape SATUSEHAT integration expects.

TODO: implement parse(ocr_json) and main(); thresholds live at the top of
the file so the review-queue tradeoff is easy to demo.
"""

raise NotImplementedError("phase 2")