Using the --defer flag for models does not consistently namespace the ml model location

[switzer]

Our development process uses the --defer flag so that we can reference production tables if they are not in the developer namespace in BQ. A sample command to build a model is as follows:

dbt run --select my_ml_model --defer --state prod-manifest --target dev --vars "{start_date: '2024-08-30', end_date: '2024-08-30'}"
10:18:28  Running with dbt=1.8.3
10:18:28  Registered adapter: bigquery=1.8.2
10:18:28  Unable to do partial parsing because saved manifest not found. Starting full parse.
10:18:35  [WARNING]: Found patch for macro "test_accepted_values_extended" which was not found
10:18:38  Found xxx models, x snapshots, xx analyses, xxx data tests, xx seeds, xx operations, xx sources, xx exposures, xxxx macros
10:18:38  
10:18:39  
10:18:39  Running 2 on-run-start hooks
10:18:40  1 of 2 START hook: ozone.on-run-start.0 ........................................ [RUN]
10:18:40  1 of 2 OK hook: ozone.on-run-start.0 ........................................... [OK in 0.00s]
10:18:40  2 of 2 START hook: ozone.on-run-start.1 ........................................ [RUN]
10:18:40  2 of 2 OK hook: ozone.on-run-start.1 ........................................... [OK in 0.00s]
10:18:40  
10:18:40  Concurrency: 3 threads (target='dev')
10:18:40  
10:18:40  1 of 1 START sql model model dbt_dev_ml.my_ml_model ......... [RUN]
10:24:41  1 of 1 OK created sql model model dbt_dev_ml.my_ml_model .... [None (2.2 GiB processed) in 30.37s]
10:24:41  
10:24:41  Finished running 1 model model, 2 project hooks in 0 hours 2 minutes and 2.87 seconds (36.87s).

Note that this model was built in the dbt_dev_ml space, which is (I think) correct.

When referencing the model, e.g. building a downstream dbt model which uses dbt_ml.predict(ref('my_ml_model'), 'source_data'), I get the following error:

 Runtime Error in model my_model_prediction (models/transformations/my_model_prediction.sql)
  404 Not found: Dataset analytics:prod_ml was not found in location EU; reason: notFound, message: Not found: Dataset analytics:prod_ml was not found in location EU

It seems like --defer does not check the dev namespace to see if a model is there first, before checking the production namespace.

Interestingly, if I compile both models (e.g. the ML model and the downstream prediction model) in the same dbt run command, then it works.