feat(engine): surface is_sleeping as a GET control endpoint mirroring /is_paused

[SuperMarioYL]

Summary

The Sleep/Wake-Up data-plane is wired end-to-end (engine.is_sleeping →
AsyncLLM.is_sleeping → IsSleepingReq round-trip through the scheduler control
client) and is GPU-tested in test_sleep_wakeup_gpu.py, but the only status
route mounted on the HTTP control surface is /is_paused. The symmetric
/is_sleeping route was missing from both the in-engine RL-control router
(vllm_compat_http.py) and the sidecar control_server.py, so callers could
poll pause status over HTTP but not sleep status.

This adds the matching /is_sleeping endpoint, mirroring /is_paused
end-to-end:

• WeightTransferManager.is_sleeping() — async accessor delegating to
  AsyncLLM.is_sleeping() (mirrors the sync is_paused() accessor at
  manager.py:94; carries await because AsyncLLM.is_sleeping is async,
  whereas AsyncLLM.weight_transfer_admission_paused is sync).
• @router.get("/is_sleeping") on vllm_compat_http.py — direct handler
  returning {"is_sleeping": bool}, mirroring the /is_paused handler.
• @app.get("/is_sleeping") on control_server.py — thin sidecar proxy to
  the in-engine RL-control app, mirroring the /is_paused proxy.

The change is purely additive: no existing route or control-plane behavior is
modified. The read semantics are point-in-time status polling, identical to
/is_paused, so the same transient-state reasoning applies to both — this is
parity with the existing route, not a new contract.

Test Plan

• TestRlControlProxy in test/runtime/test_control_server.py: spins up a
  mock RL-control app and asserts the sidecar's /is_sleeping and /is_paused
  routes relay {"is_sleeping": bool} / {"is_paused": bool} byte-faithfully.
  A missing /is_sleeping mount would 404 and fail the assertion, so this is
  a real regression guard for the new route plus parity coverage for /is_paused
  (which had no proxy test before).
• ruff check and isort --check-only clean on all four changed files;
  python -m py_compile clean.
• The control-server test suite imports the tokenspeed runtime, which pulls
  in triton / torch / tokenspeed_kernel, so it runs in GPU CI rather than
  on a non-GPU host; CI will exercise the new test.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6eb9be8d1c

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]

Document the new sleep-status endpoint

This exposes a new public GET /is_sleeping control API, but the commit does not update any user-facing documentation, leaving operators without a documented response contract or guidance on how this status differs from /is_paused. Add the route, response shape, and sleep-state semantics to the control-plane documentation as required for changed code.

AGENTS.md reference: AGENTS.md:L11-L14

Useful? React with 👍 / 👎.