Optimizer: compare mip_assign against sketch-bench #129 MILP (one-off script) #768

Change

After mip_assign lands, compare it against sketch-bench#129's MILP with a one-off script on shared-subset workloads (CPU-only weights, merge-only families, aligned scrape interval, same cost table). Record results in the PR description.

Check specifically:

• Query memory definition: Re-organize code in repo to be more OSS friendly #129 uses card × mem_bytes_per_instance (one instance per group); ASAPQuery uses n × groups × mem (Merge) / 2× (Subtract).
• Objective values and chosen deployments on each workload.

Part of #753.

🤖 Generated with Claude Code