[Alexander-Auer]

test: unit tests for the fake's hooks and the model rules, and coverage measurement

Follows #28 (merged). Tests and tooling only; no library code changes.

What is added

The fake backend (test_fake_backend.py), for what the conformance suite cannot reach without the fake's test hooks:

• refused copies keep their indexes; an injected failure is reported per job by every per-job operation; an unavailable resource fails the whole operation;
• what a resource cannot do (URL inputs or outputs, container images) is refused at submit, before anything is stored;
• a copy's environment wins over the spec's, and job IDs stay unique across submissions;
• kill aborts an active job and leaves a finished one alone;
• output can be collected again, is refused while the job runs, and an escaping member writes nothing at all;
• a purged job is gone for every operation; diagnostics read a running job's file by name.

Two of these read the fake's private job table, because the fake has no public way to show which environment a copy got or that nothing was stored.

Models, configuration and smoke tests:

• JobHandle.key escapes separators, so different handles never share a key;
• the JobStatus values stay stable, since stored statuses are compared by value;
• a configuration that is not a mapping is refused with a clear message;
• the version falls back when the package is not installed.

Tooling

• pytest-cov in the testing extras.

• [tool.coverage]: branch coverage, subprocesses measured too, protocol ... bodies and TYPE_CHECKING blocks excluded.

  pixi run -e py313 pytest --cov

Coverage of src/intercede is 99%; the two uncovered lines cannot be reached through the API.

109 tests pass on Python 3.11 and 3.13, and pre-commit passes.