Finite execution qualification
Freeze the 19 named standard-library controls in
analysis/tests/test_ilvis0_control_shards.py: eight storage/charge controls,
eight mocked process/status controls and three real-process lifecycle controls.
This tests execution machinery, not inertial observations or model fitting.
The three real-process controls each permit four child workers. They check known JSON values, different completion order, independent process sessions, inherited leases, a coordinator exception, rejection of a duplicate launch, and continuation without repeating a charged identity. Every child explicitly asserts that NumPy and SciPy were not imported. Other controls test changed inputs/results, duplicate assignments, altered deadlines, interrupted charges, incomplete transactions and cooperative failure stopping using constructed fixtures. Deliberately corrupted copies are negative tests, not modified scientific evidence.
Limits: 19 charged control starts, 12 actual child-worker spawns, 120 cooperative seconds, and at most four worker slots with one thread each. Actual native integrations, original-file reads, observed fits and protected ESA evaluations are all zero. Fixture ledgers contain simulated native/bound reservations solely to test accounting; these are not IMU calculations. Preserve every fixture and process log in the qualification output. Charge controls and worker spawns before execution; no retries, expansions or process signals. Any active children must finish normally before the next control begins. Passing does not establish one-thread numerical compatibility; the separate native qualification remains mandatory.
The qualification runs only after the current numerical allocation is free. Freeze sources, test identities and environment before launch. A failure stops remaining controls and remains visible. Rendering saved outcomes performs no native work.