Linux regression and coverage

The regression workflow builds and tests all 17 profiles on every pull request to dev or master. Its check name is auto_tests / run_tests. Workflow success alone does not enforce merging: repository rules must separately require that check. The staged Otterdog ruleset must remain disabled until this workflow is available on dev.

Run scripts/install.sh on Ubuntu 24.04, then use GCC 14 and its matching gcov with CMake and Ninja:

TX_COVERAGE=ON ./scripts/build.sh all
TX_COVERAGE=ON ./scripts/test.sh all
python3 test/cmake/usbx/test_runner.py

The installer installs gcovr 8.6 in a virtual environment and exports its location through GitHub Actions environment files. Outside Actions, activate the reported virtual environment before running coverage. CC and GCOV can select matching compiler and coverage executables. The default build and test concurrency is four. CTest runs each case once; failing cases are not retried into a passing result.

dependencies.txt records immutable ThreadX, FileX and NetX Duo revisions. FileX PR #106 is open at the pinned revision; this setup does not assume it has merged. Existing dependency checkouts must be clean and match their pins. Remove an old checkout explicitly when updating a pin. The ThreadX bootstrap supplies JUnit output without runtime edits to dependency files.

Pass profile names instead of all to either script, or select them in the manual workflow. Build and test output names the exact selection. JUnit files, test logs and build/results.txt report actual counts and elapsed test times. The generic profile's single CTest case is a build-only placeholder, not a functional USB test.

Every selected profile is instrumented when TX_COVERAGE=ON, including the generic build. Raw JSON, Cobertura XML and HTML use one inclusion policy: all compiled C sources matching common/*/src/*.c. Host controllers, simulators, optional classes and untested source lines remain in the denominator. Dependency and test code are outside that policy. No profile-specific diagnostic exclusion affects the union.

Each input must have nonempty JSON, XML and HTML, measured source lines, and repository-relative paths. The merger checks that JSON and XML agree and that its measured and covered source-line sets equal the exact unions of its inputs. Reports and execution counters are cleared before each test selection. A subset cannot consume an earlier full run's reports or publish complete coverage. Available diagnostics and coverage survive failing tests; missing coverage itself fails the runner. Full-suite line and branch floors are in coverage-floors.json. Standalone polling produces more than six billion positive hits in a single profile. Direct GCC 14 output confirms balanced loop counts. Collection uses a 2^40 suspicious-hit threshold instead of gcovr's default 2^32 heuristic; negative counts and parse errors remain fatal, and no lines are ignored to obtain a report.

Only successful complete runs on master can deploy Pages. Callers grant the permissions required for GitHub to validate the pinned reusable workflow's separate deployment job; its test job retains narrower permissions. The template bounds installation, build and test steps and uploads diagnostics on failure. The manual selector is parsed from the event file, never interpolated as shell code. All pull requests retain the full matrix; there is no conditional routing.

Dependabot targets dev once its configuration is present on the default branch. The cross-repository workflow/bootstrap revision requires a separate review and is excluded from routine action updates.