v1.2.0

v1.2.0

New Features

• Statistical confusable gate — bigram ratio detection at priority 24 (within fast-path cutoff), achieving 78% confusable detection rate with zero ML models
• MLP classifier for confusable/compound detection (F1=0.900) with ONNX export
• CMS multi-signal scoring for confusable detection (ngram + collocation thresholds)
• Neural reranker v2 — 3-gate reranking with MLM logit wiring and 19-feature MLP
• Mandatory compound expansion from 63 → 3,315 via template mining
• Title/suffix compound detection layer
• POS-based V+particle detection for broken compounds
• 28 confusable pair benchmark sentences added

Improvements

• Error budget relaxed from per-sentence skip to heavy-error-only guard
• CMS threshold reduction extended to curated confusable pairs
• Dual MLP/LightGBM training pipelines with configurable MLM logit wiring
• Benchmark confidence gap flag for analysis
• Pipeline hardening to reduce FPR on expanded benchmark
• 34 code review issues resolved across 24 files

Stats

• Benchmark: 1,138 sentences (28 confusable)
• Confusable detection: 18/23 (78%)
• Overall TP: 506 (+38 from v1.1.0)
• 4,758 tests, 75% coverage