test: make Python 3.13 async suite pass and cover it in CI

[krystophny]

Summary

Fix Python 3.13 test execution and add real 3.13 coverage to the project test jobs.

Changes

• standardize async tests on anyio with the asyncio backend
• replace Python 3.13-fragile asyncio.get_event_loop().run_until_complete(...) with asyncio.run(...)
• remove dead asyncio_mode = "auto" config (pytest-asyncio reference after its removal)
• add anyio>=4.0 as explicit dev dependency
• add Python 3.13 to the Ubuntu test matrix and the Apple Silicon MLX job

Status

• refreshed onto current upstream main (b4fa030) on 2026-04-09
• no logic changes beyond the base refresh

Validation

• python -m pytest tests/ -q --timeout=120 -> 1020 passed, 2 failed, 9 skipped, 20 deselected
• the two failures are throughput assertions in tests/test_batching_deterministic.py::TestBatchingPerformance::test_batched_faster_than_sequential[asyncio] and tests/test_continuous_batching.py::TestContinuousBatchingIntegration::test_batching_improves_throughput[asyncio]
• the async-harness changes themselves are exercising correctly; these failures are performance thresholds, not event-loop/plugin failures

[Thump604]

@waybarrios, @krystophny: brief endorsement.

This addresses issue #225 (also yours). The change is sensible: standardize on anyio with the asyncio backend, replace asyncio.get_event_loop().run_until_complete(...) with asyncio.run(...), drop the dead asyncio_mode = "auto" config, add anyio>=4.0 as a dev dep, and add Python 3.13 to both the Ubuntu test matrix and the Apple Silicon MLX job.

Validation shows 129 passed, 5 deselected. Mergeable on current main. Low risk for an infrastructure improvement that would catch Python 3.13 regressions before they ship.

[krystophny]

Force-pushed a refresh onto current upstream main (b4fa030). No logic change beyond the base refresh. Validation on current upstream plus local stack: python -m pytest tests/ -q --timeout=120 -> 1020 passed, 2 failed, 9 skipped, 20 deselected; the two failures are throughput threshold assertions in tests/test_batching_deterministic.py and tests/test_continuous_batching.py, not async/plugin failures.

[Thump604]

Refresh confirmed on head 810fb0f against upstream main b4fa030. The async-harness changes are mechanical and low-risk: standardize on anyio with the asyncio backend, replace asyncio.get_event_loop().run_until_complete(...) with asyncio.run(...), drop the dead asyncio_mode = "auto" config (pytest-asyncio was already removed), add anyio>=4.0 as an explicit dev dependency, and add Python 3.13 to the Ubuntu test matrix plus the Apple Silicon MLX job.

CI is green on every matrix entry including the new test-matrix (3.13) and test-apple-silicon (3.13) jobs. The 2 failures from the local pytest run (test_batching_deterministic.py::TestBatchingPerformance::test_batched_faster_than_sequential and test_continuous_batching.py::TestContinuousBatchingIntegration::test_batching_improves_throughput) are throughput-threshold assertions, not event loop or plugin failures, so they do not block the async-harness change class.

Approved.

[Thump604]

Rechecked on April 11: this still looks like an approved candidate, but it is currently CONFLICTING, so admin merge will not help. It needs a rebase/restack onto current main before it can move.

[Thump604]

I split the pushable/tested Python-side cleanup out into #288. The remaining Python 3.13 CI matrix update still makes sense, but I could not publish that branch from this auth context because GitHub rejects OAuth-app pushes that modify .github/workflows/ci.yml without workflow scope. So the technical state is: tested code-side cleanup is ready in #288; the workflow-file half needs someone with workflow-capable auth to carry it.