[GREEN] feat(oracle): Add unified training pipeline for deterministic corpus merge (Refs GH-123)

Adds unified_training module that merges all data sources deterministically:
- Synthetic corpus (configurable sample count, default 12,000)
- Depyler corpus (hand-crafted from tickets)
- Verificar corpus (extracted from verificar tool)
- OIP GitHub corpus (mined from Git commit history)
- Real errors file (optional)

Key features:
- Hash-based deduplication (normalized error messages)
- Deterministic shuffle using LCG with configurable seed
- Optional class balancing with max samples per class
- Comprehensive merge statistics

Tests: 6 new tests (all passing)

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

Author: noahgift

crates/depyler-oracle/src/lib.rs

@@ -33,6 +33,7 @@ pub mod synthetic;
 pub mod tfidf;
 pub mod training;
 pub mod tuning;
+pub mod unified_training;
 pub mod verificar_integration;
 
 pub use autofixer::{AutoFixer, FixContext, FixResult, TransformRule};
@@ -71,6 +72,12 @@ pub use github_corpus::{
     analyze_corpus, get_moe_samples_from_oip, CorpusStats,
 };
 
+// Unified training pipeline
+pub use unified_training::{
+    build_unified_corpus, build_default_unified_corpus, build_unified_corpus_with_oip,
+    print_merge_stats, UnifiedTrainingConfig, UnifiedTrainingResult, MergeStats,
+};
+
 /// Error types for the oracle.
 #[derive(Debug, thiserror::Error)]
 pub enum OracleError {