perf: single-pass two-phase discovery with combined filtering

[thomhurst]

Summary

• Merge ExpandSourcesForDependencies into CollectTestsWithTwoPhaseDiscoveryAsync to eliminate double enumeration of test descriptors
• Previously: descriptors were enumerated once to find dependency class names, then enumerated again for filtering and materialization
• Now: all sources are enumerated once with type-level and descriptor-level filtering applied during the same pass, while indexing everything for dependency resolution
• Net result: -140 lines, +60 lines — simpler and faster

Test plan

• Verify two-phase discovery with filters still works correctly
• Verify cross-class dependency resolution still includes dependency targets
• Verify transitive dependency expansion works
• Run full engine test suite

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

This is a clean performance optimization that successfully eliminates double enumeration of test descriptors by merging two passes into one. The refactoring:

• Removes ~80 lines of code while maintaining all functionality
• Aligns with the "Performance First" principle from CLAUDE.md
• Properly guards the cast to ITestDescriptorSource with the check on line 51
• Maintains correct dependency resolution by indexing all descriptors regardless of filter matches
• Correctly applies both type-level and descriptor-level filtering in a single pass

The logic is sound and the changes improve both performance and code maintainability.