Fixes: push every 20min, trim to last 10 entries, pull on startup for restart recovery

Fix: metrics push (summary not commit_message), crash resilience + watchdog

Bonsai-demo: multi-GPU llama.cpp server with live dashboard and metrics