Stabilize pipeline RL throughput controls

[FurtherAI]

Summary

• make stale rollout pressure the first-class autotuner signal and aggregate queue/stale metrics over complete decision windows
• keep minimum batches conservative: 85% target floor, 10% trainer-idle raise threshold, two-window collection-gain rollback, and the existing vLLM pressure gate
• report CP-compacted Megatron packed-token throughput instead of counting configured rows that CP never dispatches
• omit MoE routing payloads from trajectory Parquet and make pipeline shutdown event-driven

Why

Small minimum batches can reinforce trainer underfeed and stale-rollout pressure. The old padding-normalized underfeed denominator also overstated underfeed under context parallelism, while the packed-token numerator counted rows CP did not execute. Separately, shutdown could block forever trying to inject a sentinel into a full output queue, and legacy Parquet serialization retained large route payloads.

Validation

• full prek run --all-files
• ty across all changed production modules
• 54 focused pipeline/autotuner/Parquet tests passed, 2 skipped
• explicit coverage for vLLM pressure boundaries, two-window rollback, full-queue and blocked-source shutdown, finite-source drain, and route omission from legacy and exchange payloads

Workload-specific XML parsing, eval cadence, run naming, and parity-only test changes are intentionally excluded.