[oliverholworthy]

Run Data Designer's native check_models() before multimodal generation, so a bad credential, unavailable model, or unreachable endpoint for a later judge fails before earlier stages incur generation costs.

This runs automatically for each distinct active generator/judge configuration and any hosted summary embedding model. Shared deployments are checked once; generation batches skip repeat probes. Failed checks leave a new run unfrozen, and completed resumes still verify artifacts without credentials or model calls.

The checks make real text/embedding requests using the configured inference options. They verify basic provider readiness, not image support or judge schemas. Data Designer's DATA_DESIGNER_SKIP_MODEL_HEALTH_CHECKS=1 opt-out is respected.

Validation: UV_LINK_MODE=symlink make all — 942 passed, 2 optional Nemotron-consumer tests skipped; lint, plugin validation, generated-file checks, and docs passed. Regression coverage exercises the native readiness path with mocked model transport, including provider errors, routing, deduplication, embeddings, and resume integrity. No live model requests were used in validation.