fix(miles): keep health checks in admission mode

[TianyeGGBond]

Context

This is a narrow follow-up to the M11 router admission path.

In admission mode, an empty enabled_workers set is a valid runtime state: the scheduler may have disabled/offloaded every registered worker, and generation requests should suspend until a worker is re-admitted.

Before this change, _health_check_loop used if self.enabled_workers to decide whether admission mode was active. That makes an empty enabled set ambiguous:

• legacy/pre-admission mode: no enabled set has been declared yet, so probing the full registered worker set is correct;
• admission/zero-active mode: all workers are intentionally disabled, so falling back to the full registered set incorrectly probes disabled workers.

That fallback can fight the sleep/offload lifecycle because disabled workers are parked and should not be health-probed until they are re-admitted.

Change

• Use _admission_declared as the admission-mode switch in _health_check_loop.
• In admission mode, probe only enabled, non-dead workers, even when the enabled set is empty.
• Preserve legacy behavior before admission has ever been declared: probe registered workers minus dead workers.
• Add a focused router admission test for the zero-active case, verifying a disabled worker is not health-probed.
• Keep the production code intentionally small: no new helper and no extra feature-specific comments.

Validation

Passed:

python -m pytest tests/test_partial_sleep_wake.py::TestRouterAdmissionLifecycle -q
# 4 passed, 1 warning

Passed:

git diff --check origin/zhenyu/m11-mvp-test..HEAD

Also tried the whole file:

python -m pytest tests/test_partial_sleep_wake.py -q

Local result is limited by unrelated missing heavy dependencies in this Windows environment:

• ModuleNotFoundError: No module named 'ray' for TestEngineInfoStateMachine
• ModuleNotFoundError: No module named 'torch' for TestSchedulerPreemptClassification