A timed-out test now stops waiting

A test that declares a Timeout promises to be over by then. Until now that promise was only half kept: when the timeout fired, xunit recorded the failure, but the test's body was still parked in whatever it had been waiting on — a mesh stream that never emitted, a poll against a portal that never came up. The runner carried that blocked body along to the next test, and a wedge in one place surfaced as a slow or hung shard somewhere else.

The platform's test packages moved to the xunit.v3 4.x line (xunit.v3, xunit.v3.assert, xunit.v3.extensibility.core and xunit.v3.runner.console to 4.0.1; the VSTest adapter had already moved to 4.0.0). Its analyzer treats a test that declares a Timeout without consuming TestContext.Current.CancellationToken as a build error, and the repository builds with warnings as errors — so every one of those tests, 707 of them across 260 files, now passes the token into the call it waits on:

• observable.Await(TestContext.Current.CancellationToken) — the one sanctioned way to wait for an observable already took a token; it is now passed everywhere a test has a deadline.
• The reactive assertion chain — .Should().Within(t).Emit(…), .Match(…), .Be(…), .Complete(…) and .NotEmit(…) — gained a trailing cancellationToken parameter, threaded down to the wait itself. A cancelled wait is reported as a cancellation, never folded into "it did not emit" or, for NotEmit, into a pass the test never finished.
• Helpers a test waits on (Seed, InstallPackage, an E2E fixture's login and readiness polls) take the token and hand it to their own waits.

A synchronous test with a Timeout has nothing to hand the token to; it observes the token at its first statement, which is the analyzer's documented form for that shape.

Where the token stops, and why that is said out loud. The analyzer is satisfied by one reference anywhere in a test method, which is not the same as the body actually stopping. Every wait a test reaches directly now takes the token, and so do the helpers a test parks in longest — the E2E suites' PollAsync loops (11 files, 24 call sites), their compile- and activity-status polls, and the sync suites' ConfigWhen / QueryShows / DeliverGreenBuild. What is deliberately left is recorded rather than implied: a wait a test parks on purpose (the park is what the test measures) and a fixture's startup wait, which runs before any test owns a token. The remaining helper-level population was measured rather than waved at, and the companion entry closes it: 185 more helper-level waits now take the token, and the five that still do not each say why where they sit. What neither entry fixes — and neither claims to — is the 1,292 waits inside test bodies that are analyzer-compliant through a reference elsewhere in the method.

Nothing was suppressed to get here: neither rule was added to NoWarn, and no test's timeout, assertion or wait was changed. One fixture had two public constructors, which xunit 4.x refuses; the shape-driven one is now internal to the assembly that calls it.

One consequence is worth knowing before it surprises someone: the 4.x line ships Microsoft.Testing.Platform v2, which refuses the VSTest target on the .NET 10 SDK, so dotnet test now runs through Microsoft.Testing.Platform (opted into by a root global.json). Project paths, -c, --no-build and --filter "FullyQualifiedName~…" are unchanged; -l:trx, --logger, --blame-hang-* and -warnaserror are gone, and a local run no longer drops a .trx unless it is asked for. CI is untouched — its shards launch each project's native xunit host directly and never call dotnet test. Writing Tests carries the new flags.