feat(runtime): add Bulwark failover replay support

[hutm]

Stacked PR in the GMS-managed KV review train.

Base: review/gms-v2-latehost-05-trtllm-v2-integration
Head: review/gms-v2-latehost-06-runtime-failover-replay

Refs #10454 #10450

This PR is intentionally scoped to one commit in the train so reviewers can inspect the GMS KV ownership, engine integration, Bulwark, host-tier, router, and docs work incrementally.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[dynamo-ops]

A non-migratable error frame with no data leaves current_stream_has_output false, so the following EOF is treated as an empty stream and can replay the request after already sending an error to the client. Fix: mark the current stream as having emitted a response frame before returning from track_response for data == None, or add a separate response-frame flag and use it for the empty-stream retry check.

🤖 AI Fix

In lib/llm/src/migration.rs, update RetryManager::track_response so it sets self.current_stream_has_output = true before matching response.data.as_ref(), ensuring error/annotation frames prevent the later empty-stream replay path from recreating the request after a frame was already returned.

[hutm]

Historical review record preserved. GitHub cannot retarget this merged stacked PR; its current replacement is #11049 (position 1/18). No historical comments or reviews are being deleted.