test: fail closed on empty unit selections and expand hosted coverage

[Graffioh]

Problem and behavior

Hosted Python CI currently runs only the DS4 benchmark tests. The shared C++ runner also lets an empty filter return skip (77), lets an empty binary succeed, and lets --exact match a tag belonging to another test. CTest can therefore report success without executing the intended case.

This change makes pyproject.toml the model-free pytest entry point and runs the existing harness/concurrency/helper suites in CI. It makes empty execution/discovery fail and restricts --exact to full test names, while preserving ordinary substring/tag selection and explicit hardware skips.

A shared pytest fixture compiles small C++ binaries without CUDA/HIP and exercises the real runner and CMake/CTest adapter. The tests cover assertion failures, exceptions, missing selections, tag/name collisions, explicit skips, and empty discovery. harness/tests/README.md documents scope and tool requirements.

Validation

• 94 tests and 20 subtests passed with the CI-style pytest -q command on Linux.
• The new contract suite against main's original runner/discovery: 8 failed, 8 passed; all 16 pass with this change.
• ruff check . and git diff --check passed.
• Full server/GPU builds and model-backed qualification were not run locally; inference code is unchanged.

Based on main and independent of #636 and #687. Their qualification policies and report formats are not changed here.