Affected tests: follow NuGet package references; narrow Directory.Packages.props changes

[csa7mdm]

Closes known gap 3 of the roadmap: dotnet_test_affected's project fallback now follows NuGet package references, not just ProjectReferences.

Behavior

• Package edges (only add, never remove): a solution project's package id comes from a literal <PackageId>, the nearest Directory.Build.props, or its assembly name. When a test project's restored obj/project.assets.json references that id, it now counts as depending on that project. Test projects that were never restored are named in the note.

• Narrowing for Directory.Packages.props: the tool runs only the test projects that use the packages whose version changed. Every condition must hold, otherwise it runs the whole solution and the note says why:

  • it is the only changed file;
  • only Version attributes changed (comments and whitespace aside);
  • every project has a readable assets file;
  • git can show the previous version.

  Packages used only by source projects, such as analyzers with PrivateAssets=all, are traced through the project graph. Helper projects with no tests are never selected.

• Also fixed (a bug since 0.3.3): a changed .txt or image inside a project folder, such as TestData/expected.txt, used to be ignored and run nothing. It now runs that project's tests. .md files, docs outside every project, and a solution-root project owning docs/ don't count.

Review

Three Opus rounds.

• Round 1: 1 blocker (the narrowing could replace other selections), 4 majors and 5 minors. All fixed.
• Round 2: 1 major (an unreadable assets file read as "uses no packages") and 5 minors. Fixed by the lead session.
• Round 3: approved; 3 minors fixed.

Tests

172 in the solution; Testing.Tests went from 32 to 62. Every review scenario was written as a test and shown failing on the pre-fix code first.

Benchmark gate

The gate FAILED by its own rule. This is why it's not a regression.

• Method: both builds replayed Polly's last 40 source commits sequentially with maxSelectionSeconds: 120, then compared selected test names on commits that completed in both runs.
• Completion (22 of 40, rule needs 30): 17 commits don't complete within 120 s on either build. The threshold was set without knowing Polly well; that's my mistake, not the branch's.
• One difference, pre-existing: 8ba491d7 (RandomUtil.cs) gave 121 tests on main and 118 here. In fresh sessions, main gives 118 every time too. It's a bug in main: seeding from one TFM variant of a file with #if branches. It's fixed in its own PR (fix/tfm-variant-seeding), which should merge first.
• The other 21 commits are identical: 7,487 selected tests compared.
• Code touched: this branch doesn't modify FindAsync. Selection timing is unchanged (median 0.8 s vs 0.7 s).

🤖 Generated with Claude Code