Harden pipeline publication, trajectories, and trainer telemetry

[bradhilton]

Summary

This combines and supersedes #807 so the dependent Caladan change has one ART merge and one final SHA to pin.

• restore bounded nonstreaming tau-bench policy calls after high-concurrency streaming saturated the driver
• keep token-lineage and text-equivalent suffix reconciliation linear
• move TrainerRank LoRA conversion/serialization off the strict rank queue after immutable distributed CPU capture
• keep src/art/megatron byte-identical to main; ART PipelineTrainer continues using its original Megatron LoRA publication path
• let art.trajectory_group() accept resolved trajectories, exceptions, and arbitrary awaitables while resolving repeated awaitable identities once and preserving metadata/metrics/logs
• emit structured trainer phase and torch.compile telemetry, deduplicating plan identities on low-cardinality topology/request-mix fields while retaining exact token counts in event payloads

Tau-bench request bound

URL-backed policy clients use a 10-minute read timeout per attempt and one SDK retry, for an approximately 20-minute worst-case request budget. They no longer send private streaming/progress headers or reconstruct streamed policy completions.

PipelineTrainer isolation

The final net diff under src/art/megatron is zero. New staged export ownership lives in art.trainer_rank._lora_export and calls existing low-level Megatron helpers without changing the PipelineTrainer entry points or implementation.

Validation

• fresh blocker-only Codex review: two blockers found, fixed, and re-reviewed with no remaining blockers
• fresh Claude Fable 5 blocker review: no merge blockers
• 162 focused trajectory/history/tau-bench/telemetry tests passed
• 47 ART PipelineTrainer tests passed
• 11 legacy and staged LoRA export tests passed
• Ruff, format, ty, lock consistency, and diff checks passed
• GitHub CPU and 2xH200 checks will run on this consolidated head

Downstream

Caladan #113 must pin the actual merged SHA from this PR; it must not retain the prior #807 draft SHA.