[PatrickRitchie]

CI on a PR took 40–70 minutes, and the docs build plus its route check about 30. Almost all of that was overhead rather than tests: the unit tests themselves run in under a minute.

What changed

CI test time (ci: cut test wall-clock…)

• Coverlet's branch analysis on two generated files in JSON-cppagent (JsonEvents.g.cs / JsonSamples.g.cs, ~5,000-line constructors) took ~9.5 minutes per run. They're excluded from coverage in both runsettings files; the full unit suite with coverage now runs in under a minute locally.
• Coverage is collected on the Ubuntu leg only; Windows runs the tests without it.
• Each test project writes its own TRX file (LogFilePrefix) instead of overwriting one shared file, and a shared tools/ci/check-trx.sh guard fails the job on a missing, unreadable or empty result.
• Workflow concurrency cancels superseded PR runs; NuGet packages are cached.
• The E2E route-check shards only build the Docs-Tests project instead of the whole solution.
• docs.yml: PR runs get their own concurrency group; the Pages group now applies only to the deploy job.

Test conventions (test: unify test project conventions)

• New tests/Directory.Build.props: one version of Test.Sdk, NUnit, the NUnit adapter and coverlet for every test project (there were three Test.Sdk versions, two NUnit and three adapter versions), and a default RootNamespace of MTConnect.Tests.<Area>.
• Test namespaces normalized to MTConnect.Tests.<Area> (SHDR-Tests no longer uses the product namespace MTConnect.Shdr); the Compliance embedded-schema resource prefix follows.
• SHDR's 27 if … Assert.Pass(); else Assert.Fail(); blocks are now Assert.That(actual, Is.EqualTo(expected)), so failures show both values.
• A "Writing tests" section in CONTRIBUTING.md documents the conventions.

Compliance (test(compliance): pin the cppagent parity image…)

• The cppagent parity tests ran mtconnect/agent:latest, which now resolves to 2.8.0.4 and fails the fixture's 2.7.x version check. Pinned to the 2.7.0.7 digest.

Docs site (docs: cut the site build and route check…)

• Every API page rendered the whole ~2,500-entry API sidebar and inlined the site config, so the build wrote 3.2 GB of HTML. The API sidebar is now one per namespace and the site data is a shared chunk (metaChunk). Local VitePress build: 267 s → 117 s; HTML 3.28 GB → 0.18 GB.
• The route check walks every narrative page and a sample of the API pages (overview, each namespace page, one type per namespace): 390 of 2,445 routes, ~33 s. ROUTE_CHECK_ALL_API=true walks all of them.
• The route check no longer leaks a vitepress preview server on Windows, and its rebuild check no longer fails when dist/ already exists.
• Adds the XML doc comments missing from two durable-buffer tests. The docs build compiles every project with warnings as errors, so these broke npm run build.

How to verify

• CI: the Ubuntu leg should finish in minutes and still publish coverage; the Windows leg runs without coverage.
• Locally: ./tools/test.sh, or dotnet test tests/MTConnect.NET-Docs-Tests --filter Category=E2E after npm run build in docs/ (needs docfx).
• Verified locally: the solution builds with 0 errors; 5,929 unit tests (14 projects, CI filter) and 13 integration tests pass; Compliance passes 247/247 including the Docker parity tests; Docs-Tests passes including the E2E route check.

Known failures not addressed here

These also fail on master and are left for a separate fix:

• ShdrNetworkTests (Adapter_AnswersPing, QueueAdapter_SendsEveryQueuedValueInOrder_AndAssetsToo) are flaky: they can connect before the adapter's listener is bound.
• DeviceFilesTests.NewFile_IsCreatedFromTheTemplate fails on Linux because DeviceFiles.cs mixes a case-insensitive lookup with a case-sensitive File.Exists.
• The format job flags AgentClientConnectionListener.cs and DeviceDocumentTests.cs.

🤖 Generated with Claude Code