Read roles declared with the Roles constructor form

[woksin]

Fixed

• Roles declared with the [Roles("...")] constructor form are read by the proxy generator, not only the named-argument form (Proxy generator drops roles declared via [Roles("...")] constructor form #2381)

[woksin]

Reviewer context — deliberately kept out of the description.

Provenance. This is one commit extracted from #2417 (43ac8e46), rebased onto main and made to compile. It is the only part of that PR not already implemented in #2418 — the Screenplay fixes there (#2395, #2397, #2398, #2400) are all covered by #2418, with specs and measurements against a real application.

It did not compile as written. #2417's reviewer note says it was authored in an environment without a .NET SDK and relies on CI. Cherry-picked onto main, the fixture failed with CS0104: 'Authorize' is an ambiguous reference — Cratis.Arc.Authorization and Microsoft.AspNetCore.Authorization both declare an Authorize attribute and both were imported. Aliasing it then failed with CS1614 (ambiguous between the attribute and its Attribute-suffixed form), so the ASP.NET Core attribute is now written out in full where the fixture needs its named-argument form. A third failure followed in Release only — RCS1181, comment must be a documentation comment — which is the usual reminder that Release is the real gate here, since analyzers are errors only there.

The production change itself (MethodInfoExtensions.cs, reading ConstructorArguments as well as NamedArguments) was correct as written and needed no change.

This is agent-authored code that no human has read. Do not merge without explicit review.

Verification

• ProxyGenerator.csproj Release: 0 errors, 0 warnings.
• ProxyGenerator.Specs Release: the four new when_extracting_roles specs pass on net8.0 and net9.0 — constructor form with one role, with several roles, the named-argument form, and both forms combined.

Note on #2417. Four of its five commits duplicate work already in #2418, in overlapping files, so the two cannot both merge cleanly. Its #2397 release note also overstates the effect: on a real application that recovers nothing, because every WithMessage there resolves to a resx designer property with no constant value — which is why #2409 exists and is implemented in #2418.