fix: corrective classifier precision — reduce false positive rate

[asvarnon]

Problem

The corrective classifier matches ~45% of passages in validation (target: 10–25%). The negation-proximity heuristic (negation marker within 5 words of a triggering term + non-question filter) is too permissive for conversational transcript data.

Negation words (not, don't, shouldn't, etc.) appear naturally at high density in English conversation. Finding one within 5 words of any recurring term is statistically likely in any substantive passage — this produces false positives rather than genuine corrective patterns.

Evidence

Validation Run	Passages	Avg Length	Corrective %
Post Pre-Filter (8 entries, 309 chars avg)	1,919	309 chars	19.1%
Post-#49 (10 entries, 3,037 chars avg)	120	3,037 chars	42.5%
Post-Passage Cap (10 entries, 413 chars avg)	745	413 chars	45.0%

The 19% → 45% jump correlates with 2 new large conversational entries, not passage length. The passage cap (PR #67) fixed decisive precision but had no effect on corrective.

Possible Approaches

1. Tighter proximity window — Reduce from 5 words to 3. Risk: false negatives on legitimate patterns like "you should NOT use unwrap in library crates" (negation 3-4 words from term).

2. Require correction-specific context — Not just negation near a term, but negation + correction pattern ("don't use", "should not", "never do", "stop using", "avoid"). This would filter out incidental negation ("this is not the only way", "not just for testing").

3. Negation-term directionality — Require the negation to precede the term (most corrections say "don't do X", not "X is not"). Would reduce matches where negation follows the term incidentally.

4. Sentence-level scope — Only check negation proximity within the same sentence as the triggering term, not across sentence boundaries within the passage.

5. Combined approach — Sentence-scoped proximity (option 4) + correction-pattern requirement (option 2). Most restrictive but highest precision.

Priority

Low — does not block any current implementation. Should be addressed after the importance detection pipeline (#50) is complete and validated. The corrective category weight (1.5) is provisional until precision is acceptable.

Dependencies

• Requires feat: classify extracted passages into importance categories #49 (classification) ✅ merged
• Should be validated after any fix using the same validation script methodology

Acceptance Criteria

• Corrective precision <25% on the current 10-entry validation dataset
• No false negatives on known corrective patterns (test cases from feat: classify extracted passages into importance categories #49)
• All existing corrective tests pass or are updated with rationale
• Validation run documented in status page