Test Explorer hierarchy, command-line ceiling batching, and Solution Explorer load feedback

[MelbourneDeveloper]

TLDR

Give the Test Explorer a real Assembly → Namespace → Class → Test tree, stop large selections from killing a run with spawn ENAMETOOLONG, and put load feedback in the Solution Explorer instead of a blank view.

Details

Test Explorer hierarchy (test-discovery.ts, testing.ts)

• TestListing gains byAssembly: readonly TestAssemblyListing[] — names attributed to the assembly that contributed them. namesFrom now runs --ListFullyQualifiedTests once per assembly, because a multi-assembly listing combines names with no attribution, which is exactly what the tree's Assembly level needs. listFqns is deleted; batchAssemblies stays exported as the documented command-line-ceiling contract.
• makeAssemblyItem splits each FQN into namespace / class / test and builds the subtree. Group nodes get canResolveChildren; leaves keep the FQN as id.
• collectTests walks group nodes down to leaves, so ▶ on a class or namespace runs its members, and skips rows carrying item.error.
• pruneResults now keys off leafIdSet — group nodes never hold results.
• Discovery failure is no longer silent: makeErrorItem renders a Test discovery failed row carrying the real dotnet diagnostic (MSB1011 ambiguity, compiler errors, missing target) plus the remedy. applyDiscovery keeps a standing tree only when the sweep produced nothing at all.

Command-line ceiling (test-filter.ts, test-execution.ts, test-debug.ts, dotnet-process.ts)

• New MAX_FILTER_ARG_CHARS = 24_000 and filterBatches(). 816 selected tests reached spawn as one ~73 000-character argument and threw spawn ENAMETOOLONG synchronously, before any 'error' event could fire.
• RunInvocation.tests becomes filterIds: readonly string[]; [] means "run everything". runInto splits into batches, runs them sequentially and merges — each batch keeps its own unfiltered TRX recovery attempt, and ⏹ stops starting new batches.
• filterIdsFor passes no filter when the request is ▶-on-root, so the common case stays one invocation.
• Debug runs are never batched: every VSTest host under VSTEST_HOST_DEBUG=1 waits for an attach, so a second batch would strand its tests. An over-budget debug selection runs unfiltered and logs why.
• runDotnet wraps spawn in try/catch and settles the synchronous throw as a failed run, honouring its "never rejects" contract.

Lens resolution through the grouped tree (test-tree.ts, test-lens.ts, testing.ts)

• New test-tree.ts owns the leaf walk. runTestByMethodName matched only the TOP level of the controller, which the hierarchy turned into assemblies — so every "Run Test" / "Debug Test" lens reported a discovered test as No discovered test matching … and ran nothing. findTestByMethodName descends to the leaves and resolves neither a group node nor a discovery-error row.
• The leaf walk now has ONE implementation shared by the lens resolver and pruneResults, replacing the copies each had grown.

Solution Explorer load feedback (state.ts, tree.ts, extension.ts)

• New LoadPhase signal (idle / discovering / loading) with beginDiscovery / beginLoading / endLoadPhase. endLoadPhase is identity-checked, so a superseded scan never clears a newer phase, and both call sites in extension.ts clear it in finally — a throw out of the workspace scan or the explorer load would otherwise leave the tree spinning forever.
• rebuildTree renders a loading~spin feedback node while a phase is active; selectAndLoadSolution wraps discovery + load in window.withProgress on the view.

Other

• nuget-browser/css.ts: .material-symbols-outlined was missing font-family, so every icon rendered as ligature text.
• .github/workflows/*: dependabot pins — actions/checkout v7.0.1, codeql-action v4.37.8, setup-dotnet v6.0.0, azure/login v3.0.1, rust-cache. Cargo.lock refresh (tree-sitter 0.26.12 et al).
• src/website: js-yaml overrides pinned to patched 3.15.2 / 4.3.2; 0.19 release blog post in en/ja/zh plus its hero image.
• .claude/ → .agents/ skill relocation.
• Deleted src/website/tools/trends-results.json — a ~7 MB capture of one trends-research.mjs run's raw Google Trends API responses, read by nothing and reproducible on demand. Now gitignored.

How Do The Automated Tests Prove It Works?

src/test/suite/test-explorer-e2e.test.ts (+958/−79) drives the real VS Code Testing view against the committed fixture solution:

• tree roots are the test ASSEMBLIES — never a flat dump of test names
• under each assembly, tests group into NAMESPACE nodes split from their FQNs
• under each namespace, tests group into CLASS nodes holding exactly their tests
• every TEST is a leaf at depth 4 carrying its FQN as id and its method as label
• the tree is structurally pure: 4 levels, nothing deeper, nothing misplaced
• ▶ on a CLASS group runs exactly the tests it contains
• ▶ on a NAMESPACE group runs every test beneath it
• a discovery target that is not on disk surfaces a USEFUL error instead of silent emptiness — replaces the old assertion that it merely "yields no items and never throws"
• a folder with several projects and NO loaded solution explains itself instead of going blank
• a project that fails to BUILD surfaces the compiler error in the tree, and ▶ fabricates nothing
• a Run/Debug lens resolves a test NESTED under its assembly, namespace and class — walks the real discovered tree and asserts every fixture FQN resolves to its leaf, and that no group label (XunitCs, XunitFs, CalculatorTests) resolves to a test. Verified against the bug: with the old top-level-only walk it fails with the lens for "addsTwoNumbers" must resolve the nested test Fs.Xunit.Fixtures.addsTwoNumbers.

test-explorer-windows.test.ts (+92/−23) covers the ceiling on win32 — the platform whose 32 767-character limit produced the bug. test-explorer-frameworks.test.ts, -outcomes, -reactive and -cancellation were updated from flat-tree to hierarchy assertions (⏹ still yields no cached outcome across batches). New tree-config-e2e.test.ts (+94) covers the Solution Explorer tree, state.ts signals and config.ts round-trips end to end. nuget-browser.test.ts asserts the font-family declaration is present in the generated CSS.

Every gate in ci-build.yml was run locally on win32 before pushing: cargo fmt --check, cargo clippy --all-targets -- -D warnings, make _lint-zed, dotnet csharpier check, make _lint-dotnet (-warnaserror), prettier --check, make _lint-vsix (ESLint + tsc), both dotnet pack smoke tests, make _build-rust, make _build-dotnet, and make _test-zed including its coverage gate — all green.