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

[zhenyulincs]

Verified before approving:

• The ambiguity is real and the new predicate is the established one. The health loop was the last remaining site keying "admission active?" off if self.enabled_workers: (set truthiness), which cannot distinguish "admission never declared" (legacy — probe the full registry) from "admission declared, scheduler disabled every worker" (the legitimate F3 zero-active suspension — probe nothing). _admission_declared is exactly the flag the other two sites already use (miles_admission_disabled classification and _candidate_set()), set by all four lifecycle internals, never set on legacy paths — so legacy behavior is preserved verbatim.
• Provenance check: the replaced conditional was introduced by the port itself (iter-6 legacy-compat commit f1f1e2f); iter 7 (1978e31) introduced _admission_declared and converted the other call sites but missed this one. This PR completes that conversion — no upstream-original miles code is modified, the upstream health-loop body is untouched.
• Blast radius confirmed: on the buggy path, parked/offloaded workers get probed and can be marked dead, causing _notify_workers_changed churn and false "Marking as DEAD" logs; _enable_worker_internal clears dead + failure counts, so it self-heals on re-admission — state pollution and noise rather than permanent quarantine. Which is also why runtime smokes never surfaced it.
• Ran the test file on this head: the TestRouterAdmissionLifecycle class passes 4/4 including the new zero-active probe test (the other classes in the file fail only on missing torch/numpy in my slim env — unrelated to this change).

One non-blocking simplification for a follow-up (not this PR): the new inline branch duplicates _candidate_set() exactly (declared → enabled - dead, else → registry - dead), and _candidate_set's docstring already claims the health loop uses it. urls = list(self._candidate_set()) would remove the third copy of this set arithmetic and make that docstring true — and would have prevented exactly this kind of missed-conversion bug.

[JunzheJoe]

LGTM.