Fix silent 0.0 return on number parse failure in extract_numbers.py

[Pran-Ker]

Summary

• normalize_number() returned 0.0 when float() conversion failed, silently converting invalid input into a meaningful financial value
• In financial contexts, zero and "unparseable" are very different — a false zero could create incorrect inconsistency groupings in find_inconsistencies() or mask data quality issues
• Now returns None on parse failure; unparseable entries are skipped entirely from the extraction pipeline

Changes

• normalize_number(): Returns None instead of 0.0 on ValueError, with updated return type Optional[float]
• NumberInstance.normalized: Type updated to Optional[float] to reflect that values may not always be parseable
• extract_numbers(): Skips entries where normalized is None so invalid data never enters the pipeline
• find_inconsistencies(): Added None guard on the division used for percentage-diff grouping

Test plan

• Run extract_numbers.py on a presentation with valid numbers — output unchanged
• Run on content with malformed number strings — entries are skipped instead of appearing as 0.0
• Run with --check flag — inconsistency detection still works correctly
• Verify no regressions in number grouping logic

🤖 Generated with Claude Code