[moetemp]

What changed?

This PR adds @nexus.error, which maps the cases of an error enum to exception classes.

/// @nexus.error details="stream-error-details" field="reason" name="stream-error"
///   python-extends="temporalio.exceptions.ApplicationError"
enum stream-error-reason {
  /// @nexus.fallback
  stream-error-reason-unspecified,
  /// @nexus.error
  stream-error-reason-not-found,
  /// @nexus.error retryable=true
  stream-error-reason-storage-unavailable,
}

• Each marked case gets a class under one base. Another case, or one that this code doesn't know, raises the base class. An unknown number decodes as the @nexus.fallback case of feat: decode an unknown proto enum number as a marked fallback case temporalio/nexgen#321.
• A class is non-retryable unless it has retryable=true. Its failure type is the class name, and its details hold the details record, so a Local Activity failure maps to the same class.
• Python's calls convert a failure to the mapped class, in Workflow code and in a client. Go and Java ignore the directive with a warning.

TypeScript and .NET get the classes and the function that picks one. Their calls convert a failure in a follow-up PR.

Why?

The proposal raises one error class per Store reason, such as StreamNotFoundError. The facade maps the reasons by hand today.

How did you test it?

Link to a test plan if any -

• Unit Tests
• Staging
• End to End Tests

Python tests raise the mapped class from a client call and from a Local Activity failure. Python and TypeScript tests decode a reason that this code doesn't know from the wire, and get the base class.

[moetemp]

Closing this. The Stream error classes move into sdk-python by hand, since StreamErrorDetails is a proto that the SDK already decodes. nexgen doesn't need @nexus.error.