v1.4.0

v1.4.0 — Meta-Classifier & Semantic Confusable Detection

Highlights

• Meta-classifier post-filter: A learned logistic regression model (41 features including one-hot error type, word frequency, and context signals) replaces manual per-strategy confidence thresholds. FPR dropped from 34.5% to 18.6%.
• ConfusableSemanticStrategy (priority 48): MLM-enhanced confusable detection using masked language model logit comparison with asymmetric thresholds.
• Rich Suggestion objects: Suggestion class with confidence and source metadata, backward-compatible (inherits from str).
• Per-request CheckOptions: Runtime overrides for context_checking, grammar_checking, max_suggestions, and use_semantic.

Added

• Kinzi (င်္) and consonant stacking variant support in confusable candidate generation.
• MLM post-filter to suppress invalid_word and dangling_word false positives.
• Expanded confusable pairs from 87 to 124+ with 9 linguistics-audit additions.
• Expanded colloquial variants from 83 to 91 entries; removed 20 standard modern Burmese words incorrectly classified as colloquial.
• Homophone morphological guard expanded to 2 prefixes and 4 compound suffixes.
• Error.severity property with computed severity based on action type.
• Candidate fusion enabled by default.

Changed

• Config split: algorithm_configs.py split into 4 focused modules (algorithm_configs.py, text_configs.py, strategy_configs.py, infra_configs.py). All existing imports continue to work.
• Benchmark consolidated to 1,146 sentences with 18 duplicate IDs fixed.
• Confidence gates expanded to 15 error types.
• Zero-TP detectors disabled, weak detectors heavily gated.
• Mutex/override infrastructure fully removed.

Fixed

• MLM threshold comparison now operates in logit space (was incorrectly using probability space).
• context_checking=False in CheckOptions preserves word validation instead of skipping all validation.
• Suggestion objects properly coerced to str for Cython compatibility in normalize functions.
• Streaming checker no longer dumps tracebacks to stderr for handled exceptions.
• Word segmentation degrades gracefully when resources are unavailable (offline/CI environments).
• Resource download timeouts wrapped as clean TokenizationError instead of unhandled exceptions.
• Long verb chains reassembled correctly, dangling particle scan capped.
• Invalid_word and dangling_word FPs reduced via post-validation compound splitting.