chore(deps): WolverineFx 6, last 9.x packages, xunit v3 + xUnit1051 adoption

[ChrisonSimtian]

Finishes the dependency sweep — every remaining major — plus the xunit v3 migration. Two commits, each self-consistent and independently buildable.

1. WolverineFx 6 + the last 9.x packages (880fd06)

• WolverineFx + WolverineFx.Postgresql 5.18.0 → 6.22.0
• Microsoft.Extensions.ServiceDiscovery 9.3.1 → 10.8.0
• Microsoft.Extensions.Http.Resilience 9.4.0 → 10.8.0

Wolverine 6 compiled with zero source changes and the entire unit suite stayed green — while the fleet was completely broken. Two separate breaking changes, both runtime-only:

a) Core no longer ships the runtime Roslyn compiler (GH-2876). Handler code is generated at runtime under TypeLoadMode.Dynamic, so startup died:

Wolverine is running in TypeLoadMode.Dynamic … but no IAssemblyGenerator (Roslyn) is registered.
Core WolverineFx no longer ships the runtime compiler.

Added WolverineFx.RuntimeCompilation to the two hosts that call UseWolverine (ARD + ZDF agents); it self-registers. The alternative — pre-generated code + TypeLoadMode.Static — trades a build step and a regeneration discipline for faster cold start. Reasonable for the container images later, but not something to fold into a dependency bump.

b) ServiceLocationPolicy now defaults to NotAllowed (5.x was AllowedButWarn), so codegen refused to build a handler needing container resolution. CrawlShowHandler needs it twice over: IEnumerable<IBroadcasterCrawler> is an opaque lambda registration, and IEpisodeRepository's graph reaches EF Core's own DbContextOptions factory, which we don't control.

Set to AllowedButWarn — restoring exactly the 5.x behaviour. Deliberately not Allowed, so Wolverine keeps warning and the nudge to make these inline-constructible stays visible rather than being silenced.

Verified by running it, not by compiling it

Because neither failure was visible to the compiler or the test suite:

• Migrator applied all migrations against Postgres 18.3
• ZDF agent started → scheduler dispatched CrawlShowCommand over the durable Postgres transport → handler received it → crawled the real ZDF API → Crawled 18 episode(s) for 'heute-show' on zdf
• 18 rows in Episodes; Wolverine created its 8 message-store tables
• No errors in the agent log

2. xunit v3 migration (228be5b)

All five test projects: xunit 2.9.3 → xunit.v3 3.2.2. The xunit.runner.visualstudio 3.1.5 and Test.Sdk 18.8.1 from the last PR already support v3, so no runner change was needed.

What the migration forced:

• IAsyncLifetime now returns ValueTask (was Task) and extends IAsyncDisposable — so PostgresFixture and the three repository fixtures needed new signatures.
• Test projects are executables in v3 — added <OutputType>Exe</OutputType>.
• TngTech.ArchUnitNET.xUnit had to go. It still depends on xunit.assert 2.x, which cannot coexist with v3, and 0.13.3 is the latest release. Swapped for core TngTech.ArchUnitNET plus a local ArchRuleAssert.Check() that evaluates the rule and fails through Assert.Fail, listing every violation.

The architecture tests were checked for false-passing, not assumed green

Replacing an assertion mechanism can silently turn 8 passing tests into 8 no-ops, so I inverted one rule — asserting Infrastructure does not depend on Domain, which it plainly does — and confirmed a proper failure:

Architecture rule violated: Types that reside in namespace with full name matching
"Krautwatch\.Infrastructure" should not depend on any types that reside in namespace
with full name matching "Krautwatch\.Domain" because Domain is the core of the hexagon…

Both the rule description and its because reason survive. Reverting restored 8/8. The layer rules still genuinely enforce.

3. Adopt xUnit1051 (c00725c)

xunit v3 ships analyzer rule xUnit1051 — calls accepting a CancellationToken should pass
TestContext.Current.CancellationToken so the runner can cancel tests. It does not exist in v2, so
commit 2 introduced 258 warning instances across 129 call sites. Adopted rather than suppressed:
it earns its place in Live.Tests, where a single test spends 6–16 minutes downloading from a CDN
and previously could not be cancelled at all.

Build is back to 0 warnings / 0 errors.

"Append an argument" turns out to be wrong more often than right — three distinct cases:

Case	Fix
Token slot omitted	append the token
Slot already filled with default	replace it — appending produced 106 CS1501 errors
Other optional params before the token (FindShowAsync(query, client = "ard", ct = default))	named argument, or it lands in the client slot

NSubstitute setups became Arg.Any<CancellationToken>(), matching the convention the verification
calls in this suite already used. Binding a setup to one specific token would silently stop matching
once the handler forwards a different one — and a substitute that stops matching returns null rather
than failing loudly.

Verification calls moved from default to the test's own token, e.g.
Received(1).AddAsync(Arg.Any<DownloadJob>(), TestContext.Current.CancellationToken). That is a
slightly stronger assertion than before: it now proves the token is actually propagated from the
test through the handler to the repository.

Framework calls that could not be classified automatically were done by hand — FluentValidation
ValidateAsync, HttpContent.ReadAsStreamAsync, Stream.WriteAsync, and EF's FindAsync, whose
cancellable overload takes key values as an object[], so the call shape changes rather than just
gaining an argument.

Testing

	Result
./build.sh Test	125 tests green, Live correctly excluded, 0 warnings / 0 errors
./build.sh TestLive	16/16 green against real ARD/ZDF — trait filtering works under v3 in both directions
Fleet	ZDF agent crawled real ZDF over the durable bus, 18 episodes persisted
Commit 880fd06 alone	builds clean (0 warnings) — the split is genuinely bisectable

Notes

• A live test flaked once and was confirmed transient. Downloads_a_full_Extra3_episode_from_ARD failed with a SocketError after 1m54s mid-download from ARD's CDN; re-runs passed 16/16 twice. Real full download against a live CDN, unrelated to these changes.
• Live suite wall-clock varies a lot — 6m to 16m across runs today, entirely CDN throughput. Something to bear in mind if it is ever put on a timeout.
• Commit 2's message was amended to remove an incorrect "0 warnings" claim: I had grepped for warning CS and missed the xUnit1051 analyzer warnings, which CI surfaced. It now states the warnings introduced and points at commit 3.
• tests/Domain.Tests is in the solution but contains zero test files — pre-existing, and why it never appears in test output. Either give it tests or drop it.
• Microsoft.Extensions.{ServiceDiscovery,Http.Resilience} needed no source changes despite the 9.x → 10.x jump.

🤖 Generated with Claude Code