v0.10.0 — cross-framework provider switching + llama.cpp validated

LocalHarness now speaks to more than one local inference framework and can switch between models served on different backends.

What's new

• Switch across frameworks: /model re-points the live agent between models served on different endpoints, re-probing capabilities and token counting on the way. The picker groups models by framework + host.
• llama.cpp validated as a second provider: auto-detected, exact token counting via its /tokenize, served-window discovery, and a real multi-step agentic tool-call round trip verified end to end on GPU.

Honest scope

• Ollama and LM Studio use the same OpenAI-compatible client path, but their full on-GPU certification — and the harness launching non-vLLM servers itself (lifecycle) — land in 0.11. This release is switch + attach + llama.cpp validation.
• Pre-1.0: interfaces and config schema may still change.