Deep RuVector integration for the Contrastive CSI Embedding Model:
- MicroLoRA on ProjectionHead: rank-4 LoRA adapters (1,792 params/env,
93% reduction vs full retraining) with merge/unmerge, freeze-base
training, and per-environment LoRA weight serialization
- EWC++ consolidation in Trainer: compute Fisher information after
pretraining, apply penalty during supervised fine-tuning to prevent
catastrophic forgetting of contrastive structure
- EnvironmentDetector in EmbeddingExtractor: drift-aware embedding
extraction with anomalous entry flagging in FingerprintIndex
- Hard-negative mining: HardNegativeMiner with configurable ratio and
warmup, info_nce_loss_mined() for efficient contrastive training
- RVF SEG_LORA (0x0D): named LoRA profile storage/retrieval with
add_lora_profile(), lora_profile(), lora_profiles() methods
- 12 new tests (272 total, 0 failures)
Closes Phase 7 of ADR-024. All 7 phases now complete.
Co-Authored-By: claude-flow <ruv@ruv.net>
0826438e0e