Add unit tests for schema parsing functions (parse_review_decision, parse_sanity_result, etc.)

[wittekin]

Problem

src/millstone/policy/schemas.py contains complex multi-layer fallback parsing logic (JSON → regex → heuristics) for:

• parse_review_decision
• parse_sanity_result
• parse_builder_completion
• parse_design_review

These are safety gates — a parsing bug could cause millstone to approve bad code or reject good code. The fallback paths are currently untested at the unit level; they're only exercised indirectly through integration tests.

Fix

Add tests/unit/test_schemas.py (~200 LoC) covering each parser with:

• Valid structured JSON input
• JSON embedded in prose ("Here is my review:\njson\n{...}")
• Malformed JSON that triggers regex fallback
• Completely invalid/garbage input
• Edge cases: empty string, whitespace-only, non-ASCII

This ensures the fallback chain works correctly and any changes to parsing logic are caught immediately.