test: restore the global metrics collector instead of leaking it (#3780)

[nicoloboschi]

Fixes #3780.

What was wrong

create_metrics_collector() replaces a module global in metrics.py and nothing ever put the previous collector back. NoOpMetricsCollector ignores its arguments; the real MetricsCollector inspects them (if cached_input_tokens > 0). Provider tests hand it a bare MagicMock usage object, so once anything installed the real collector, every test that ran after it in the same xdist worker was one MagicMock away from:

TypeError: '>' not supported between instances of 'MagicMock' and 'int'

Which files got hit was decided by xdist's distribution, so the failure count moved between runs of identical code — which is what made the suite useless as a signal.

The API lifespan is not the only poisoner. worker/main.py:72 installs a collector as a side effect of constructing create_worker_app, and tests/test_health_probes.py builds that app with no lifespan at all.

The fix

• metrics.py — reset_metrics_collector(collector=None), the missing counterpart to create_metrics_collector(): puts back a previously installed collector, or the no-op default.
• api/http.py — the lifespan snapshots get_metrics_collector() before startup and restores it after shutdown_tracing(). Worth doing on its own merits: an app that starts and stops otherwise leaves its own collector, holding a reference to a now-closed DB pool, installed for the rest of the process.
• tests/conftest.py — an autouse fixture doing the same around every test, following the existing _cleanup_leaked_span_recorders fail-safe. This is the part that actually fixes the suite, since it also covers the construction-time paths that never run a shutdown.

Verification

A/B with a real in-suite poisoner, no synthetic plugin — test_health_probes.py builds the worker app, test_reasoning_effort_explicit_config.py is one of the files the issue lists:

$ uv run pytest tests/test_health_probes.py tests/test_reasoning_effort_explicit_config.py -p no:randomly -n 0
  autouse fixture off: 13 failed, 12 passed
  autouse fixture on:  25 passed

tests/test_metrics_collector_lifecycle.py covers reset_metrics_collector both ways, plus a DB-free test that drives app.router.lifespan_context over a stub engine and asserts the collector is restored — and asserts the app really did install one first, so it cannot pass vacuously.

Scope

• No production fix, deliberately. _usage_from_openai_response already coerces the field with getattr(usage.prompt_tokens_details, "cached_tokens", 0) or 0, so a real provider response yields 0. Only a MagicMock reaches that comparison.
• create_worker_app still installs the collector at construction time. Harmless in production — that process runs until exit — and the autouse fixture covers the test consequence. Giving the worker app a lifespan is a restructure beyond this bug.

[strix-security]

Strix Security Review

No security issues found.

_{Updated for 1cdb243.}

---

Reviewed by Strix
_{Re-run review · Configure security review settings}