A Databricks v2 agent configured with the bare id gpt-6-astra got a 404
(the endpoint is databricks-gpt-6-astra). The harness retried the batch
10 times over ~40 minutes before posting the raw error, and the picker
could not offer the right id because the v2 gateway listing lags v1.
Dead-letter -32002 immediately with a notice that names the model,
supplement v2 discovery with v1 databricks-* endpoints, add a gpt-6
capability rule, and name the rejected model in the desktop copy.
Signed-off-by: Diem Nguyen <diem@squareup.com>
0b754688b7