Refactor _resolve_matching_subtype to use an LLM agent for ambiguous
cases instead of pure static rules:
Architecture (3 layers):
1. Static deterministic rules: M:N→MxN, 1:N (WS-MAST/TRAN-KEY),
二段階, 混合 — high confidence, no LLM needed
2. LLM agent: ambiguous cases (N:1 vs 1:1, M:N→M vs M:N→N)
- _MATCHING_SUBTYPE_AGENT_PROMPT with 5 subtypes
- Calls existing hina.hina_agent._parse_llm_response for parsing
- Minimum confidence threshold 0.4 to gate low-quality LLM output
3. Fallback: conservative defaults (M:N or 1:1) when LLM unavailable
This follows the original architecture design: agent handles the
hard classification problems that static analysis alone can't resolve.
Regression: 745 passed (unchanged).
Issues found through matching program classification analysis:
1. dedup_vs_nodedup: 0.85→0.50 for negative detection (no WS-PREV-KEY
is not strong evidence for '含まず')
2. validation_vs_keybreak: 0.80→0.55 for has_counter (counter is a
generic pattern, not specific to key-break)
3. simple_vs_two_stage: 0.80→0.50 for non-open-close-open pattern
(sequential OPEN is the default for most programs)
Result: matching programs now correctly classified:
- MT01-03/18/20 → マッチング ✅ (was 項目チェック)
- MT16-17 → 二段階マッチング ✅ (unchanged)
- MT32 → 項目チェック(重複含む) ✅ (correct: has WS-PREV-KEY)
- VL01 → 項目チェック(重複含む) ✅ (correct)
- CSV → CSV合并 ✅ (correct)
Regression: 745 passed (3 test expectation bounds updated)
M1: Cache confusion-pair confidences in Path B (eliminate redundant
resolve_confusion_pair re-calls in _path_rule_engine)
M2: Resolve contradictions in Path C instead of hardcoding
resolved_count=0 in _path_llm_assisted
M4: Add DIVIDE_25 to contradiction pair coverage (50-25, 100-25)
and update test_contradiction_pairs_defined to verify all 3 variants
NB-076
hangshuo652