Connect validated four-model RNGD producers to live PD serving #234

Problem

The four-model matrix has validated saved RNGD KV exports, but live HTTP PD still lacks integration of the validated Qwen3-4B, KONI and Gemma producers. Gemma production imports are enabled by #231; its HTTP evidence currently replays recorded exports.

Evidence

The existing #199 Qwen3 worker already implements SubprocessPrefillWorker's shared-memory/request protocol, but the service fixes Qwen3-0.6B, context512 and width<=512. #227 supplies validated model-specific producers and true-length export semantics at longer widths. Production #232 handles fresh Gemma hybrid KV, with numerical limits retained in #233.

Scope

Reuse the existing shared-memory worker contract and production scheduler/HTTP owner. Refactor model producers into callable adapters used by both measurement and live requests; support Qwen3-0.6B, Qwen3-4B, KONI-8B and Gemma3-1B with explicit compiled capacity and checkpoint geometry. Validate input boundaries, EOS, cancellation/reuse, device ownership and same-fixture native comparisons. Start with fresh full-prefix transfers; retain cache/overlap gaps explicitly under #201 instead of claiming completion.

Done when

Each model handles real live RNGD prefill -> ATOM HTTP decode in an owned allocation with recorded correctness, lifecycle and timing evidence, and runs are distinguished from saved-KV replay. Numerical mismatches and unsupported cache paths stay explicit; communication-library changes only if the existing protocol is demonstrated insufficient.

Owner

Codex; maintainer review. Related #199/#201/#227/#231/#233.