feat: Model Control Center with embedded LLM inference

[CalebisGross]

Summary

• Self-contained embedded inference: mnemonic binary runs its own LLM via llama.cpp with ROCm GPU acceleration. No LM Studio, no external servers, no API dependencies required.
• Model Control Center dashboard: new Models tab with runtime model management — hot-swap GGUF models, toggle between embedded and Gemini API, view model status and swap logs.
• SwitchableProvider: wraps embedded + API providers with runtime toggle. Switching to API unloads models from VRAM; switching back reloads them.
• ChatML compatibility: fixed prompt formatting for Qwen 3.5 (was using Felix-LM format), added /no_think to disable reasoning tokens, <think> stripping, <|im_end|> stop sequences.
• Qwen spoke fusion: export script now fuses spoke matrices for fewer GPU kernel launches, RQ4 quantizer inner-dim fix for ssm_conv1d.

Files Changed

Area	Files	What
Provider	embedded.go, switchable.go, provider.go	Hot-swap, Unload/Reload, manifest listing, SwitchableProvider, ModelManager interface
API	routes/models.go, server.go	GET/POST /api/v1/models, /api/v1/models/active
Dashboard	models.js, index.html, nav.js, app.js	Models tab, status cards, swap UI, provider toggle
Startup	runtime.go, serve.go	SwitchableProvider wiring, ModelManager injection
Training	export_qwen35_spokes.py, quantize_rq4.py	Spoke fusion, inner-dim quantization fix

Test plan

• go vet ./... clean
• go test ./... all pass
• golangci-lint run zero issues in changed packages
• API endpoints respond correctly (embedded and API modes)
• Dashboard Models tab renders, swap buttons work
• VRAM released on Gemini switch, reloaded on embedded switch
• Dreaming/consolidation/abstraction agents produce content with ChatML fix
• Full lifecycle test with fused RQ4 spokes

🤖 Generated with Claude Code