Testing and pull-request gates

Let's Infer has one canonical regression command:

python3 -m tools.core_regression

It runs the CLI, benchmark, gateway, replica-orchestration, node, repository contract, and portable macOS contract suites in isolated processes. It also rejects unregistered test modules, so adding a new directory cannot silently remove coverage from CI.

For a focused local iteration:

python3 -m tools.core_regression --list
python3 -m tools.core_regression --suite cli --suite regression

What every core pull request proves

The required Core regression suite check runs on pull requests to main and release. It verifies:

• every public and internal CLI leaf has an explicit scope and a parseable, mocked dispatch path;
• update availability, application, failure, rollback, and concurrent polling paths remain deterministic without contacting production services;
• model installation, runtime selection, lifecycle, protection, gateway, node, replica, benchmark, audit, and publication contracts pass;
• Python and installer sources parse cleanly; and
• the native Watchdog builds and passes CTest on Linux.

The release workflow uses the same Python runner. A release therefore cannot substitute a smaller test set than the one reviewed on its pull request.

Tests must use temporary directories and mocks for network, service-manager, container, hardware, and upstream update boundaries. They must never depend on a developer's installed models, API keys, node identity, Docker daemon, or GPU.