v1.4.0: RAG reliability & filter stability

What changed

• Smart RAG retrieval pipeline — multi-stage search with FTS fallback eliminates silent failures on SVD and hierarchical documents
• CMSIS-SVD file import — import peripheral register maps for AI-assisted embedded debugging
• KV cache quantization and speculative decoding — new LLM performance settings for faster inference
• Matryoshka embeddings — configurable embedding dimensions for flexible RAG accuracy vs speed trade-offs
• Snapshot-based filter restoration — toggling filters no longer loses ANSI colors or inflates line counts
• Per-session resize controller — fixes SIGWINCH storms that caused shell redraws during layout changes
• Pop-out window reliability — filter state preserved across pop-out/reattach, stale subscriptions cleaned up
• Windows >2GB GGUF support — vendored llama-cpp-sys-2 patch fixes 32-bit file offset overflow
• Qwen3.5 model support with native chat template formatting