MLSys P0-4: fleet library (6 SKUs + 8-GPU node) and the reference admission program

Part of the MLSys 2027 fleet-admission readiness set — see the tracking issue.

Why

This is where the paper's evidence comes from, and where "bug fixes here and there" actually lands. Writing real SKU declarations and a real admission program will flush const-eval and checker bugs that synthetic tests do not reach.

Scope

• fleet/{a100-40,a100-80,h100-sxm,h200,b200,mi300x}.vx, from vendor spec sheets, citing them in comments
• one 8-GPU node declaration with NVLink / PCIe / IB edges
• reference inference program: weights + KV + activations as const-generic arithmetic over (layers, heads, dim, dtype, context, batch, TP)
• file each bug found as its own issue and link it back here

The spec-sheet citations are not optional politeness — the paper's capacity numbers are only as credible as their provenance.

Estimate: ~1 week including fixes.
Depends on: fp8 element types, --machine flag.

Start by checking the fp8 element-type and --machine prerequisites, then inspect the fleet/ declarations named in the scope. Add the six vendor-cited SKU files, the 8-GPU node edges, and the reference inference program; use the const-eval and checker behavior to identify issues, and file each discovered bug separately.