feat(zip): add offline US ZIP enrichment to form-fill pipeline

[CrepuscularIRIS]

What

Adds a ZipResolver class that infers missing US ZIP codes from already-extracted form fields (city/state, embedded address text, or county strings) using an offline SQLite database — no network calls at fill time.

Why

Fixes #435

Cal Fire responders often remember a location (town, district, address) but not the ZIP code during field reporting. FireForm can now auto-fill missing ZIP fields from context already present in the transcript, reducing manual correction and speeding up report completion in low-connectivity environments.

How

A new src/zip_resolver.py module provides ZipResolver.enrich(extracted_data: dict) -> dict. It is called in src/filler.py immediately after the LLM extracts field values and before they are written positionally to the PDF.

Three inference strategies run in priority order:

1. Extract an embedded 5-digit ZIP from any address field text ("123 Main St, Sacramento, CA 95814" → 95814)
2. City + state lookup via uszipcode (handles both abbreviations "CA" and full names "California")
3. Parse a "City, ST" pattern from address or county fields ("Pine Valley, CA" → 91962)

Key design decisions:

• Non-breaking: existing non-empty ZIPs are never overwritten
• Conservative fallback: leaves field unchanged if ZIP cannot be inferred — a missing ZIP on an official report is better than a wrong one
• Graceful degradation: if uszipcode is not installed, enrichment is silently skipped via a try/except ImportError
• Immutable: returns a copy of the input dict, never mutates LLM state
• uszipcode was chosen over pgeocode for this use case: it is US-specific, bundles its data as a local SQLite DB (no GeoNames download step), and has a purpose-built city/state search API

Scope

• Included: ZipResolver class, integration into filler.py, 20 unit tests, uszipcode dependency
• NOT included: non-US postal codes, integration tests requiring a live Ollama instance, UI changes

Verification

• PYTHONPATH=src python -m pytest src/test/test_zip_resolver.py -v → 20 passed, 0 failed
• Tests placed in src/test/ to match the CI workflow (tests.yml runs src/test/)
• No changes to existing test files; no regressions
• src/filler.py change is 2 lines (import + one enrichment call); all existing behaviour preserved

Test configuration:

• Python 3.11.15
• OS: Linux (Ubuntu)
• uszipcode mocked in all tests — zero network dependency in the test suite

AI Disclosure

AI-assisted implementation, manually reviewed, tested, and verified line by line.