pytest-jev: Semantic assertions for LLM outputs by TypeSafe's Jev

Evaluates decision probabilities on labeled examples and compares accuracy, calibration and automation coverage. It helps choose when to accept Jev decisions or escalate them, and supports regression checks when model behavior changes.

Adds diagnostic and routing judgments to OpenTelemetry log records. Jev helps identify records worth further model analysis while retaining the original archive, separating triage decisions from long-form diagnosis.

Removing the abstain option forces Jev to answer unanswerable KoBBQ items, causing accuracy to drop from 0.950 to 0.000 and the stereotype rate to jump from 0.03 to 0.79.