fix(kv-router): scope KvRouter cancellation to a child token

[dagil-nvidia]

Summary

• Cherry-pick of fix(kv-router): scope KvRouter cancellation to a child token #11390 to release/1.3.0
• Scopes KvRouter cancellation to a child token so cancelling a request no longer tears down the shared router token.
• Adapted for release/1.3.0 API (no health-monitor plumbing; cancellation token threaded through subscriber/worker-query paths only).

Original PR

• Main PR: fix(kv-router): scope KvRouter cancellation to a child token #11390
• Merge commit: 88ddd8a
• Cherry-pick commit: 4b6b04a

Conflict Resolution Notes

• subscriber.rs / worker_query.rs: kept release/1.3.0 signatures; added scoped CancellationToken params instead of main-only workers_with_configs / health-monitor args.
• e2e_harness.py: omitted run_cache_salt_isolation_test (not present on release/1.3.0).

Test plan

• CI passes
• Verified fix works on release branch

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 4b6b04abbd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Keep router tasks alive through graceful shutdown

When Runtime::shutdown() is called while a KV-routed endpoint still has in-flight or queued requests, this token is cancelled in shutdown Phase 1 because Runtime::child_token() is a child of endpoint_shutdown_token, and Phase 2 only waits for graceful endpoints afterward (lib/runtime/src/runtime.rs). That tears down the scheduler/indexer/sequence tasks before the active requests finish, regressing graceful shutdown behavior; the mocker explicitly avoids this by deriving from primary_token() so it only stops in Phase 3. Use an independently cancellable child of the primary token (or equivalent) so Drop can stop this router without cancelling the runtime, while runtime shutdown keeps router support alive until Phase 3.

Useful? React with 👍 / 👎.

[dynamo-review-agent]

Adding cancellation_token to public Indexer::new breaks downstream dynamo-llm callers. Fix: keep the old public constructor and move the token-taking implementation behind a new crate-private constructor used by KvRouter.

🤖 AI Fix

In lib/llm/src/kv_router/indexer/mod.rs, restore pub async fn Indexer::new(component, kv_router_config, block_size, model_name) with the old signature and make it call a new pub(crate) async fn new_with_cancellation_token(..., cancellation_token: CancellationToken) containing the current implementation; update lib/llm/src/kv_router.rs KvRouter::new to call Indexer::new_with_cancellation_token.

[dynamo-review-agent]

Adding cancellation_token to public KvScheduler::start breaks downstream Rust callers. Fix: preserve the old public start signature and add a token-aware internal start helper for router-owned scoped cancellation.

🤖 AI Fix

In lib/llm/src/kv_router/scheduler.rs, restore pub async fn KvScheduler::start(...) without the cancellation_token parameter, have it call a new pub(crate) async fn start_with_cancellation_token(..., cancellation_token: CancellationToken) containing the current implementation, and update lib/llm/src/kv_router.rs KvRouter::new to call KvScheduler::start_with_cancellation_token.

[dynamo-review-agent]

Adding cancellation_token to public create_multi_worker_sequences breaks downstream dynamo-llm callers. Fix: keep the old public function and put the new cancellation-token plumbing in a crate-private variant.

🤖 AI Fix

In lib/llm/src/kv_router/sequence.rs, restore pub async fn create_multi_worker_sequences(...) without the cancellation_token parameter, have it call a new pub(crate) async fn create_multi_worker_sequences_with_cancellation_token(..., cancellation_token: CancellationToken) containing the current implementation, and update lib/llm/src/kv_router/scheduler.rs to call the token-aware variant.