fix(domains): GTIN float rewrite, EIDR check character, ambiguous dates with time, healthcare tz

[kevincostner17]

Summary

This PR fixes four domain-validator bugs.

• Retail GTIN (domain='retail' turns float-loaded GTIN 4012345678901 into a different GTIN-14 #229).
  • Bug: a blank GTIN cell makes the CSV column float64, so 4012345678901 was checked as "4012345678901.0". The strip_nondigits repair then rewrote it to 40123456789010, a different GTIN-14 that still passes the mod-10 check.
  • Fix: integral float cells are now read as integer text for validation and repair. The repair skips non-integral numbers and any text with a . followed by digits, so those values are flagged but never rewritten.
• EIDR check character (EIDR check character uses ISO 7064 Mod 37,2; real EIDR IDs fail MD-C002 #259).
  • Bug: eidr_check_char used pure ISO 7064 MOD 37-2 (with *). The EIDR ID Format spec uses the hybrid MOD 37,36 system over 0-9A-Z, so every real EIDR ID failed MD-C002.
  • Fix: the algorithm is replaced and * is no longer accepted.
  • Tests and docs: the known-answer test had been generated by the old algorithm and now uses ten published EIDR IDs. CONTRIBUTING_DOMAINS.md is corrected.
• Finance ambiguous dates (Finance date repair guesses MM/DD for ambiguous dates that include a time #260).
  • Bug: the FIN-003 "refuse to guess DD/MM vs MM/DD" guard only matched a bare date, so 03/04/2024 09:30 bypassed it and was parsed month-first.
  • Fix: the guard now matches the numeric date prefix whatever follows it.
• Healthcare tz handling (tz-aware vs naive datetimes crash contracts, semantic checks, domain checks and cdc_profile #233, part 2).
  • Bug: the deceased-after-birth, age-range, encounter-duration and shared ge_date checks compared offset-aware FHIR dateTime values with naive dates or mixed offsets, and raised TypeError.
  • Fix: they now parse with utc=True (naive values are read as UTC) and use pd.Timestamp.now(tz="UTC").

Behaviour changes

• EIDR: IDs whose check character came from the old MOD 37-2 code are now flagged by MD-C002. A * check character is always rejected.
• GTIN: float-loaded columns with integral values now validate instead of failing GS1-002. GTINs that lost a leading zero stay flagged and are not repaired.
• Finance dates: ambiguous dates that include a time are left unresolved instead of being read month-first.
• Date comparisons: healthcare and education checks now compare in UTC.

Tests

• tests/domains/test_retail.py:
  • a CSV with a blank GTIN is not rewritten
  • a float GTIN column is validated as integer text
  • the repair never drops a decimal point
• tests/domains/test_media.py:
  • ten published EIDR IDs validate
  • known-answer check characters
  • old MOD 37-2 check characters and * are rejected
• tests/domains/test_spec_conformance.py: the published EIDR ID in row 0 is no longer flagged (MD-C002 rows are [1, 2, 3]).
• tests/domains/test_finance.py: an ambiguous date with a time stays unresolved; an unambiguous date with a time is still repaired.
• tests/domains/test_healthcare.py:
  • a parsed FHIR Patient with an offset deceasedDateTime
  • deceased-before-birth and implausible age with offset values
  • an Encounter with mixed offsets, compared in UTC

Verification

• ruff check .: all checks passed
• mypy src/freshdata: no issues found in 202 source files
• pytest (-m "not online and not large"), py3.12 / pandas 2.3.3: 4262 passed, 6 skipped
• pytest, py3.9 / pandas 1.5.3: 4258 passed, 10 skipped
• Each issue's reproduction script fails on main and passes on this branch, on both lanes.

Closes #229
Closes #259
Closes #260
Refs #233 (part 2: healthcare validator)

[github-actions]

FreshData benchmark report — performance

• freshdata: ?
• python: ?
• platform: ?

fixture	n_rows	n_cols	p50 s	p95 s	peak MB	repair %	false-repair %	preserve %	trust	monotonic	export %

Authored-code reduction (Metric 6)