AutoMegaKernel: Command for Claude Code

Launch the unattended AutoMegaKernel (AMK) autoresearch driver on $ARGUMENTS (parse as model [gpu] [minutes|iters]; default gpu = rtx5090; default model = toy). This is the headless keep/revert campaign, point it at a (model, gpu), give it a budget, and it runs the whole correctness-gated methodology unattended, resumable and crash-proof, growing the flywheel.

HARD HONESTY RULES (state them, obey them):

• Correctness FIRST: NEVER a latency without a correctness PASS vs the CPU ReferenceVM. Keep iff correct AND >= 1% faster than the incumbent.
• validate-before-launch: an unsafe ScheduleConfig is a clean REJECTED, never a hung GPU.
• Edit surface is ScheduleConfig + kernel_knobs ONLY, never kernel code, never vm/ or the ABI.
• Measured-gpu latency is drift-robust; impossible sub-roofline latencies are withheld.
• The morning "best" is a speedup vs AMK's OWN default schedule, NOT a claim of beating cuBLAS/vLLM (AMK is within ~13% of cuBLAS at batch-1, behind it).

Do this:

1. Confirm the environment first with amk_doctor() (CLI: amk doctor), torch/cuda availability, device name, registered targets. For a real measured-gpu campaign you need a CUDA GPU; otherwise the fitness is analytic predicted (use device="cpu").
2. Launch the driver with amk_autoresearch(model, gpu, minutes=<M>, iters=<N>, device="auto", overnight=<bool>, cold=<bool>). CLI fallback: amk autoresearch <model> --gpu <gpu> --minutes <M> --device cuda [--overnight] [--cold] (or --iters <N> --device cpu for a fast deterministic run). Use overnight=true with a long minutes (e.g. 480 for ~8h): no plateau-stop, basin-hops to fresh regions while always preserving the global best, bounded memory, checkpointed every iteration (re-run the same command to continue), and writes a wake-up report.
3. It is resumable + crash-proof: a CUDA error / timeout in one iteration is logged and the run continues. Re-running the same command continues the same campaign.
4. When it finishes (or to inspect progress), report from the orchestrator: amk_orchestrate_status(), amk_orchestrate_report() (CLI: python amk_orchestrate.py status / report), and for an overnight run read workspace/amk_overnight_report.md (best config + speedup-vs-baseline + milestones + restart count).

Report: best schedule_id, best latency_us + latency_kind, pct_of_roofline, and the speedup vs AMK's own default schedule (NOT vs cuBLAS/vLLM).

Use only the canonical names: MCP amk_doctor/amk_autoresearch/amk_orchestrate_status/ amk_orchestrate_report; CLI amk doctor|autoresearch and python amk_orchestrate.py status|report.