Source-gen: generic method runs only the first [Arguments] row per inferred type #7096

Summary

In source-generated mode, a generic test method with several [Arguments] rows only runs the first row for each inferred type argument. Any later row that infers the same T is silently dropped. Reflection mode runs every row.

Found while working on #7094.

Repro

public class GenericMethodSameTypeArgumentsTests
{
    [Test]
    [Arguments(1)]
    [Arguments(2)]
    [Arguments(3)]
    public async Task SameType<T>(T value)
    {
        await Assert.That(value).IsTypeOf<int>();
    }

    [Test]
    [Arguments(1)]
    [Arguments("a")]
    [Arguments(2)]
    [Arguments("b")]
    public async Task MixedTypes<T>(T value)
    {
        await Assert.That(value is int or string).IsTrue();
    }
}

Run with --treenode-filter "/*/<namespace>/GenericMethodSameTypeArgumentsTests/*".

Expected

7 tests in both modes: SameType runs for 1, 2 and 3, and MixedTypes runs for 1, "a", 2 and "b".

Actual

Nothing fails or warns. The extra rows just never get discovered.

Root cause

TestMetadataGenerator.CollectConcreteInstantiations builds one concrete instantiation per type-argument combination and skips any combination it has already seen (processedTypeCombinations, keyed by BuildTypeKey). For [Arguments] rows, the instantiation also records the row that produced it (SpecificArgumentsAttribute), and EmitConcreteDataSources / EmitConcreteAttributeFactory then emit only that row as the entry's data source. So the second [Arguments(2)] row infers int, hits the dedup, and is discarded, and the int entry only carries [Arguments(1)].

Related observations (not part of this fix)

Generic classes have similar gaps, but there reflection mode doesn't run every row either, so the two modes already agree:

• [Arguments(1)] [Arguments(2)] class Foo<T>(T value) runs 1 test in both modes.
• class Foo<T> { [Test, Arguments(1), Arguments(2)] void M(T value) } runs 1 test in source-gen and 0 in reflection.