The failure path, in short was:

• Try and train a model but use an invalid text value for task
• Then try and train a model but use an invalid text value for test_sampling
• Then retrain the same project name but with correct configs

Bug persists if the deployed id doesn't actually appear in snapshots