baseline models

Resolve 20–30% “this week” slowdowns with a Bash-first, Linux-native baselining workflow: install hyperfine/perf/numactl/sensors/Python, fingerprint OS/kernel/CPU/microcode, lock power and thread/NUMA placement, run a real ONNX Runtime ResNet50 benchmark, track p50/p90, throughput, IPC/cache misses and thermals, and store results in git or containers to spot, explain, and prevent regressions after updates.