feat: add curated target framework sets and stop stripping Roslyn com…

[kieronlanning]

…ponent namespaces

Roslyn component names are no longer stripped from RootNamespace. SourceGenerator, SourceGenerators, SourceGeneration, Generators, Analyzers, CodeFixers and CodeFixes were in the suffix list, and because the stripped value is what reaches the compiler as build_property.RootNamespace, every file whose namespace kept the suffix failed IDE0130. That was not an edge case: it broke five of the six repositories consuming this SDK, each of which keeps the suffix because a generator, its analyzers and its code fixes are separate assemblies that each need their own namespace. A component's suffix is part of its identity rather than noise. Add one back per project with NamespaceRemoveSuffix Include if a repository really wants it collapsed.

Curated target framework sets replace a hard-coded literal as the way a project's targets are decided:

<PurviewTargetFrameworkSet>Supported</PurviewTargetFrameworkSet>

Current net10.0 (the default)
Latest net11.0
Supported net10.0;net11.0
Broad net8.0;net9.0;net10.0 every shipped release
All net8.0;net9.0;net10.0;net11.0 Broad plus the newest major

A multi-targeting repository otherwise hard-codes its TFM list in every project, so a lifecycle change is an edit in every repository that has to be found and kept consistent. 'Current' applies when nothing is selected and resolves to the same single TFM this SDK used to hard-code, so the out-of-box result is unchanged. Broad and All differ by one real product decision - whether a package commits to the newest release - so they are separate names rather than one set that silently widens every consumer.

Overrides are layered: an explicit TargetFramework/TargetFrameworks in the project always wins, then PurviewTargetFrameworks redefines a set for a repository, then the definitions above. A single-entry set resolves to TargetFramework so it does not pay for an outer build. An unrecognised name fails with PurviewInvalidTargetFrameworkSet rather than silently falling back and dropping every target asked for. Roslyn components keep netstandard2.0; one that declares a set itself is told through PurviewTargetFrameworkSetIgnored, while inheriting a repository-wide selection stays silent because that is the normal arrangement.

Fixed: an explicitly declared RootNamespace was overwritten. FixRootNamespaceTarget runs after the project body, so it could not tell an author's value from a derived one and rewrote Foo.CodeFixers to Foo - then every file in it failed IDE0130 against a namespace nobody asked for. The declaration is now recorded during evaluation and the target leaves it alone.

Fixed: project detection read commented-out declarations. The SDK is imported before the project body is evaluated, so TargetFramework, TargetFrameworks, IsRoslynComponent, IsRoslynComponentOnly and IsPackable are read from the project XML. XML comments were not stripped, so a TargetFramework left inside a comment counted as a real declaration - and for a Roslyn component that meant skipping the netstandard2.0 default, letting the generic .NET default fill in, and shipping an analyzer the IDE cannot load. Comments are stripped now. Only unconditional declarations are matched, because a Condition cannot be evaluated that early and guessing either way is worse than not matching; a declaration that exists only under a Condition is reported as PurviewConditionedProjectDeclaration. Declaring a property unconditionally and then augmenting it conditionally is correct and stays silent.

Fixed: 168 InternalsVisibleTo attributes per shipped assembly, 53 of them malformed. Three grants were emitted per test type - from AssemblyName, TargetProjectName and MSBuildProjectName - with no empty guard, and TargetProjectName has no default, so every test type produced a grant with an empty assembly name ('.UnitTests'). The three are usually the same string, so each real grant was emitted up to three times. Measured on a project using this SDK: 168 grants down to 116, malformed 53 down to 0, duplicates 0. The existing tests only asserted the attribute type appeared, which is why this went unnoticed; the generated values are asserted now.

Fixed: mirrored .agents content dirtied every consuming repository. The packaged blanket .gitignore is only written for a folder this SDK wholly owns - a skill is its own folder - so files mirrored directly into .agents/agents and .agents/prompts received none and stayed tracked, and an SDK upgrade that changed a prompt showed up as a local modification. A blanket rule there is not an option: a consuming repository keeps its own authored prompts and agents beside the mirrored ones. The sync now writes .agents/.gitignore from the manifest it already builds, listing only the mirrored paths, so authored files keep their normal status. Both that file and the packaged per-skill one ignore themselves rather than being committed - '!.gitignore' left a mirrored, never-committed file permanently untracked, and because the nearest .gitignore wins it also overrode the generated parent list.

Tests: 436 passing, up from 409. The new coverage is the part that was missing rather than an addition for its own sake - the detection, namespace and InternalsVisibleTo behaviour had no assertions on observable output, which is precisely why these defects survived.