feat: Databricks GPT-5.x endpoints unblock + re-run eval Measurement A

[deanban]

Problem
On workspace dbc-6395977d-5b1c all 5 databricks-gpt-5-* endpoints (and their short aliases like 54mini) return 403 PERMISSION_DENIED: rate limit of 0, verified against both the native serving-endpoints surface and the MLflow AI Gateway surface (https://7474647691394693.ai-gateway.cloud.databricks.com/mlflow/v1) with a Gateway-scoped PAT. This is a workspace-wide Databricks-side throttle, not token-scoped. Blocks Measurement A of the native Databricks Mosaic AI provider eval.

Proposed solution
File a Databricks support ticket (subject + workspace ID + error payload are documented in docs/runbooks/databricks-mosaic-provider.md under "GPT-5.x is blocked — support-ticket playbook"). Once the throttle lifts, re-run sema eval run on showcase/cbioportal_to_omop/slices/dev_slice_poc.yaml with LLM_MODEL=databricks-gpt-5-4-mini as a third condition and append results to eval-runs/databricks-mosaic-comparison-2026-04-22.md.

Alternatives considered
Stay on llama-3-3-70b via AI Gateway (works but ~2x slower than the OpenRouter baseline; measured in the 2026-04-22 eval report).

Reference: tasks.md §13.1; docs/runbooks/databricks-mosaic-provider.md.