[lesnik512]

Fixes #337.

HealthChecksInstrument.define_health_status always ran asyncio.gather over the framework check and every probe in health_checks_probes. With no probes, the default, that wrapped a single coroutine in a Task and cost two event-loop iterations per health request.

With no probes, it now returns await self._define_framework_health_status() directly. With probes configured it still uses gather, so they keep running concurrently; test_healthchecks_probes_run_concurrently covers that path. FastStream's _define_framework_health_status still pings its brokers through its own gather; I left it alone, since a broker list is the case gather exists for.

Test, written first and failing on main:

• test_healthchecks_status_without_probes_skips_gather: with asyncio.gather wrapped in a MagicMock that still calls through, define_health_status() returns True and the mock is never called. On main it was called once.

Measured with benchmarks/frameworks on macOS (Apple M2, CPython 3.14), in-process, per health request. Both columns use the same venv with only this file swapped, over 3 interleaved runs of 3 repeats:

scenario	main	this branch
Litestar health, over a plain handler with the same JSON body (health_ping)	+46 to +49 µs	+0.3 to +1.4 µs
FastAPI health, over bare	+58 to +59 µs	+4.8 to +5.2 µs

#337 measured +7.1 to +0.4 µs (Litestar) and +22.6 to +9.3 µs (FastAPI) on Linux. The saving is larger on a Mac because each event-loop iteration costs more there (KqueueSelector.select(0) against EpollSelector), as the issue noted. The Linux numbers are the ones to expect in production.

Checked locally: lint and ty are clean, all 400 tests pass on the branch rebased onto #348, and coverage is 100%.

Not breaking.