Reflect terminal PD producer failure in readiness and request admission #246

Problem

A terminally failed subprocess prefill producer leaves the batched service marked ready and accepting requests that cannot succeed.

Evidence

Actual tiny Torch subprocess plus HTTP server on ea09aca: /ready200 before worker death; after owned child kill and receiver termination, _failed is set but /ready still200/ready=true and the next completion returns500/non-retryable. The decoder service state remains idle. Baseline JSON preserved in research records.

Scope

Expose subprocess producer readiness, propagate configured pool-member availability, and use it consistently in batched HTTP readiness and new-request admission. Reject unavailable producer work with retryable503 before scheduling. Preserve per-request error recovery, non-PD/legacy worker compatibility, and explicit idle-boundary replacement. Do not silently reroute requests, replay work or add automatic restart.

Done when

Real subprocess/HTTP tests prove200->503 on terminal failure, no new admitted work,200after explicit healthy replacement with correct output, and per-request errors do not falsely poison readiness. Existing serving/PD tests pass.

Owner

Codex implementation; maintainer review/merge.