test(router): add GMS placement E2E coverage

[hutm]

Stacked PR in the GMS-managed KV review train.

Base: review/gms-v2-latehost-09-3-router-event-plane-gms-d2d-worker-publish
Head: review/gms-v2-latehost-09-4-router-event-plane-gms-d2d-e2e

Refs #10457 #10458 #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]

When GMS is enabled, the harness starts servers on devices 0,1 by default even for gpu_1/single-GPU tests, so those tests can fail by trying to open GPU 1 on a one-GPU runner. Fix: default the GMS device list to the resolved single test GPU unless DYNAMO_ROUTER_E2E_GMS_DEVICES is explicitly set.

🤖 AI Fix

In tests/router/e2e_harness.py, update maybe_router_gms_servers so the unset DYNAMO_ROUTER_E2E_GMS_DEVICES default is str(int(os.environ.get("DYNAMO_ROUTER_E2E_GPU_START_INDEX", "0"))) instead of "0,1", leaving explicit multi-device coverage controlled by DYNAMO_ROUTER_E2E_GMS_DEVICES.

[dynamo-ops]

The SGLang default now leaves CUDA and piecewise CUDA graphs enabled, but the same-GPU two-worker tests rely on disabling them to avoid the documented capture-time VRAM spike. Fix: keep the default disabled and allow opt-in via the env override.

🤖 AI Fix

In tests/router/test_router_e2e_with_sglang.py, change the SGLANG_ARGS["disable_cuda_graph"] default passed to _env_bool("DYNAMO_ROUTER_E2E_SGLANG_DISABLE_CUDA_GRAPH", ...) from False to True.

[grahamking]

+58,833 -1,057

I think you accidentally included all of the stacked PRs.

[grahamking]

Include everything, I think this is a mistake needs rebase maybe. 60k lines.

[hutm]

Historical review record preserved. GitHub cannot retarget this merged stacked PR; its current replacement is #11053 (position 10/18). No historical comments or reviews are being deleted.