feat(observability): Export forward pass metrics to local collectors

[nealvaidya]

Summary

• expose event-plane envelope metadata from FpmEventSubscriber through a new recv_envelope() API while preserving the existing recv() behavior
• add dynamo.common.export_forward_pass_metrics, an opt-in subscriber that discovers one or more Dynamo components and forwards decoded forward-pass metrics as NDJSON over a local Unix socket
• preserve publisher ID, event sequence, publication timestamp, component, namespace, worker ID, and DP rank for downstream integrity checks and time alignment
• add a focused serialization test for the local export record

Motivation

Benchmark-scoped collectors need the full forward-pass event history from Dynamo rather than a backend-specific vLLM connection or tracking mode's latest snapshot. Keeping discovery and decoding in Dynamo provides one backend-neutral integration point for vLLM, SGLang, and TensorRT-LLM publishers.

The local socket boundary also lets a collector run without NATS: it can use etcd discovery, the configured Dynamo event plane (for example ZMQ), and a TCP request-plane runtime solely to initialize the distributed runtime.

Compatibility

• FpmEventSubscriber.recv() remains unchanged for existing callers.
• recv() and recv_envelope() share the existing receive mode and remain mutually exclusive with start_tracking().
• The exporter is opt-in and does not change worker or frontend startup behavior.

Validation

• cargo fmt --all -- --check
• cargo check --manifest-path lib/bindings/python/Cargo.toml --lib
• Ruff format and lint checks for the exporter and test
• focused pytest for test_export_forward_pass_metrics.py

[datadog-official]

⚠️ Warnings

🚦 2 Pipeline jobs failed

PR | backend-status-check     

PR | dynamo-status-check     

ℹ️ Info

🎯 Code Coverage (details)
• Patch Coverage: 31.58%
• Overall Coverage: 40.38% (-6.67%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: d9a419f | Docs | Give us feedback!}

[nealvaidya]

Closing as superseded by #11110, which persists forward-pass metrics at the producer as rotating JSONL gzip trace segments. The downstream SRT-Slurm and Tachometer integrations now consume that official trace format, so Dynamo no longer needs the in-process subscriber/exporter added by this draft.