M1 — Jetson 4096-token BFCL Agent Harness

Current focus: keep GenieClaw fast, reliable, and measurable on NVIDIA Jetson Orin Nano 8GB with a 4096-token local context. In scope: BFCL quick-router and local-LLM tool-call scoring, high-signal home/family memory fixtures, deterministic device state, typed-tool routing, memory retrieval accuracy, compact prompt/tool budgeting, Jetson/aarch64 CI, and Jetson hardware validation when possible. Out of scope: broad prompt growth, generic chatbot/provider churn, UI/product/community/hardware work, toy demos, untested native/runtime changes, and PRs that make the agent less native, slower, less deterministic, or harder to test.