Validate four-model live GPU PD HTTP over long inputs and lifecycle events #240

Problem

The four-model coverage has standalone GPU HTTP and offline RNGD-to-GPU continuation, but lacks live GPU PD HTTP evidence over long inputs and concurrent clients.

Evidence

The production serve_batched factory already supports a separate GPU prefill device. The HTTP matrix validator does not expose that option or verify remote prefill counters.

Hypothesis and scope

Exercise actual GPU0-prefill/GPU1-decode for Qwen3-0.6B, Qwen3-4B, KONI-8B and Gemma3-1B using the existing HTTP matrix (128/2048/8192 inputs, queued clients, cancellation/disconnect, invalid requests, cache conditions where supported). Add only CLI wiring and evidence checks to the validator. Preserve numerical differences and failures. Timings from this lifecycle harness are observational, not a controlled native-vs-PD speedup claim.

Done when

Four-model outcomes and exact commands/versions are recorded, actual remote work is verified by counters, discovered defects are fixed or linked to explicit investigations, and limitations including cross-host RNGD-to-GPU live transport remain visible.

Owner

Codex research implementation; maintainer review/merge.