v1.3.0

v1.3.0 — Candidate Fusion Pipeline

Highlights

• Candidate fusion pipeline: Calibrated Noisy-OR fusion replaces the single-winner selection. All validation strategies can now fire at every position, and a confidence-based arbiter determines which errors to surface — significantly reducing false positives.
• Confidence gates: Word error suppression and context evidence guards filter out low-confidence results on structurally-clean text.
• Pipeline conflict resolution: Override matrix and candidate arbiter for resolving cross-strategy conflicts.

Added

• StatisticalConfusableStrategy (priority 24): Bidirectional bigram ratio for detecting confusable word pairs in context.
• ConfusableCompoundClassifierStrategy (priority 47): MLP-based compound word detection using ONNX.
• Data-driven calibration: Per-strategy breakpoints and reliability weights loaded from YAML.
• Error.source_strategy field tracks which strategy produced each error.
• SpellCheckerBuilder.with_candidate_fusion() for fluent configuration.
• Expanded confusable pairs from 37 to 87, mined from production database.
• Expanded grammar rules: additional auxiliary verbs, null-copula fix, grammar frequency guard.

Changed

• ConfigPresets.accurate() now enables candidate fusion by default.
• NERConfig split into lightweight module for faster imports.
• Lowered default strategy confidences and confusable thresholds for better recall.
• BrokenCompound span extraction uses word positions directly instead of sentence.find().
• N-gram strategy applies error correction to all context words.

Fixed

• ထင်ရှား no longer incorrectly flagged as aspiration error.
• ကုမ်ပဏီ now correctly suggests ကုမ္ပဏီ instead of ကုန်ပဏီ.
• Removed disputed ha-htoe correction for နူနာ per MLC dictionary.
• CalibrationData loader handles invalid thresholds gracefully.
• Narrowed overly broad exception catches in classifier and suggestion pipeline.
• Neural reranker warns on missing stats file instead of silently using unnormalized features.