fix(runtime): give event publishers unique discovery identities

[zhongdaor-nv]

Summary

• Give every EventPublisher a fork-safe, per-incarnation publisher ID and use the same ID in both EventEnvelope and event-channel discovery.
• Keep the existing four-part discovery key and serialized DiscoveryInstance::EventChannel shape while allowing multiple publishers for the same process/component/topic.
• Track ZMQ publisher connections by full DiscoveryInstanceId and ignore unrelated topics.
• Keep asynchronous discovery unregister work alive through graceful-shutdown Phase 2.
• Make discovery registration terminology unambiguous: conversion uses a documented default_instance_id, registration code logs only the final instance_id, and EventPublisher logs use publisher_id. The old public with_instance_id method remains as a compatibility alias.
• Add direct ZMQ regression coverage for two same-topic publishers and drop-to-immediate-recreate, including a fresh identity/sequence space, eventual discovery state, and continued delivery from both surviving and recreated publishers.
• Remove the mocker E2E autouse fixture that forced NATS, so nondurable mocker coverage exercises the default ZMQ event plane.

The root cause was that each publisher bound its own ZMQ endpoint and owned its own sequence counter, but all publishers in one process registered with the process-level discovery ID. Their discovery keys therefore collided, subscribers only connected to one endpoint, and dropping any colliding publisher could remove the shared registration.

This also makes (publisher_id, sequence) valid for broker deduplication across multiple publishers and prevents a late unregister from an old publisher incarnation from removing its replacement. The wire schema and discovery path structure are unchanged. The public Rust DiscoverySpec::EventChannel variant now requires publisher_id.

Validation

• Rebased onto origin/main at 2469510c7d6 and resolved the event-plane conflict while retaining both main's cached NATS subject path and this change's publisher lifecycle handling.
• cargo fmt --all
• git diff --check
• cargo test -p dynamo-runtime transports::event_plane:: --lib — 18 passed
• Drop-to-recreate direct-ZMQ regression repeated 20 times without failure
• cargo test -p dynamo-runtime --lib — 433 passed, 2 ignored
• cargo clippy -p dynamo-runtime --lib --tests -- -D warnings
• cargo check -p dynamo-llm -p dynamo-backend-common
• maturin develop --uv
• pytest -q -s 'tests/router/test_router_e2e_with_mockers.py::test_router_decisions[nats_core-tcp]' — 1 passed in 40.75s with the default ZMQ event plane, two workers, and dp_size=4
• pytest -q -s 'tests/router/test_router_e2e_with_mockers.py::test_router_decisions_disagg_round_robin_prefill_dp_rank[no_bootstrap-prefill_first]' — 1 passed in 80.35s with the default ZMQ event plane
• pytest -q -s 'tests/router/test_router_e2e_with_mockers.py::test_router_decisions[jetstream-tcp]' — 1 passed in 40.64s, confirming durable coverage still opts into NATS
• pre-commit run --files on all changed Rust and Python files

Summary by CodeRabbit

• New Features

  • Event channel registrations now use a distinct publisher identity, improving how multiple publishers on the same topic are tracked.
  • Subscribers now connect only to matching ZMQ topics, reducing unrelated connections.

• Bug Fixes

  • Fixed unregister and logging behavior so the correct instance is reported and removed.
  • Improved publisher shutdown handling so one publisher can be stopped without affecting others.

[datadog-official]

⚠️ Warnings

🚦 1 Pipeline job failed

Docs link check | lychee     

ℹ️ Info

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 42.21% (-9.87%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 89aa35e | Docs | Give us feedback!}

[devin-ai-integration]

Devin Review found 1 potential issue.

[coderabbitai]

Walkthrough

The PR splits discovery ownership from event-channel publisher identity, threads the new publisher id through event publisher registration and ZMQ subscriber matching, updates discovery logging, adds shutdown-safe unregistration, expands ZMQ integration coverage, and removes a router mocker environment override.

Changes

Discovery and event-plane identity flow

Layer / File(s)	Summary
Discovery identity contract lib/runtime/src/discovery/mod.rs	DiscoverySpec::EventChannel now carries publisher_id, and with_instance_id distinguishes owner ids from publisher ids when building DiscoveryInstance values.
Discovery registration ids and logs lib/runtime/src/discovery/kube.rs, lib/runtime/src/discovery/kv_store.rs	Kube and KV-store discovery registrations now capture owner and registered instance ids separately, and unregister logging uses the instance being removed.
EventPublisher identity and shutdown lib/runtime/src/transports/event_plane/mod.rs	EventPublisher generates a random publisher_id, includes it in NATS and ZMQ direct discovery specs, and registers a shutdown guard during async unregistration.
ZMQ subscriber matching lib/runtime/src/transports/event_plane/dynamic_subscriber.rs	DynamicSubscriber::start_zmq tracks active endpoints by DiscoveryInstanceId, filters add/remove events by topic, and the extractor tests cover topic and transport matching.
ZMQ multi-publisher test lib/runtime/src/transports/event_plane/mod.rs	A Tokio test covers two ZMQ publishers on the same topic, independent discovery and delivery, and unregistering one publisher while the other remains active.

Router mocker harness

Layer / File(s)	Summary
Remove NATS pinning fixture tests/router/test_router_e2e_with_mockers.py	The mocker e2e router fixture that forced DYN_EVENT_PLANE to nats for etcd-backed tests was removed.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

🚥 Pre-merge checks | ✅ 3 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 75.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.
Description check	⚠️ Warning	The PR description is detailed, but it omits the required template sections, especially Where should reviewer start? and Related Issues.	Add the missing template sections and include a related issue reference or the no-issue checkbox, plus a reviewer start point.

✅ Passed checks (3 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly captures the main runtime change: assigning unique discovery identities to event publishers.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}