@spec collect(term(), term(), keyword()) ::
  {:ok, Imp.Prediction.t()} | {:error, term()}

Runs one program call through stream/3 and returns what the program returned: {:ok, %Imp.Prediction{}} or {:error, reason}, the same as Imp.call/2. Every output field is on the prediction.

With provider_stream: true the prediction is the one the stream ends with, whatever :include_final_prediction says. Without it, the program runs once and its prediction is returned without chunking it. If the stream emits an error chunk, collection stops and returns {:error, reason}.

Parses provider chunks into incremental typed field updates.

The parser follows the ChatAdapter delimiter format: [[ ## field ## ]] followed by field text. A field is emitted when the next field delimiter arrives or when the stream ends.

Streams one program call as an Enumerable of chunks.

With provider_stream: true, Imp executes the real program and streams from its named predictors as they are reached. :stream_listeners can select intermediate output fields; without listeners, normalized provider events from every named predictor are yielded. The final typed prediction is yielded by default and can be disabled with include_final_prediction: false.

Composed modules expose their predictor names through the ordinary optimizer predictor callback in Imp.Module. A program with no named predictors returns a terminal {:provider_stream_unsupported, module} error. Without provider_stream: true, the program runs once and the result is chunked locally (grapheme by grapheme, or through the :chunker function).

Either way a failure is the last element of the stream, as %Imp.Streaming.Messages.StreamResponse{chunk: {:error, reason}, done: true}.