fix(router): disclose selected worker in non-KV modes for 1.3.0

[ishandhanani]

Backports #11245 to release/1.3.0 using the release branch's existing session-affinity API. It discloses the resolved worker for non-KV routing only after dispatch succeeds, preventing failed attempts from poisoning retry metadata.

CLOSES: DYN-3353

Summary

• Add prepared-dispatch paths that return the worker selected after transport resolution.
• Record worker attribution for standard, session-affinity, direct, and LoRA routing only after a response stream is created successfully.
• Preserve occupancy accounting when transport fallback selects a different worker.

How This Was Implemented

• The source cherry-pick conflicted only in session_affinity/push_router.rs: release/1.3.0 retains the older prepare(request, worker_id, dp_rank) callback, while main uses the AffinityTarget callback introduced by feat(router): coordinate session affinity across replicas [DYN-3249] #11079 after the release branch split.
• The resolution keeps the release callback and rank behavior unchanged, while adapting fix(router): disclose selected worker in non-KV modes #11245's post-success recording order to it.
• revert(router): keep session affinity local [DYN-3249] #11217 is not included because the release branch never received feat(router): coordinate session affinity across replicas [DYN-3249] #11079 and already has the local-affinity behavior that revert(router): keep session affinity local [DYN-3249] #11217 restores.

Walkthrough

Mental model

sequenceDiagram
    participant Router
    participant PushRouter
    participant Backend
    participant Tracker
    Router->>PushRouter: Select and prepare request
    PushRouter->>Backend: Dispatch to resolved worker
    Backend-->>PushRouter: Response stream created
    PushRouter->>Tracker: Record resolved worker and rank

Loading

State and ordering

The tracker is removed from the prepared request and returned as dispatch metadata. Worker attribution is committed only after dispatch returns a stream; preparation or dispatch failures leave the tracker unchanged so retry and fallback paths can record the worker that actually succeeds.

Boundaries and limitations

This backport does not import the distributed-affinity work or its later revert. The public release callback remains (worker_id, dp_rank), and existing release routing behavior remains unchanged.

Validation

• cargo test -p dynamo-runtime -p dynamo-llm push_router::tests --lib (41 passed)
• cargo clippy -p dynamo-runtime -p dynamo-llm --lib -- -D warnings
• cargo fmt --all -- --check
• git diff --check

[datadog-official]

⚠️ Warnings

🚦 2 Pipeline jobs failed

Docs link check | lychee     

PR Reminder Full CI Comment Bot | pr_reminder     

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: c554435 | Docs | Give us feedback!}

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.