RestateModelWrapper should not wrap `SdkInternalBaseException` as generic Exception; classify non-retryable provider errors as terminal

[prescod]

When a journaled model step fails, Restate correctly uses SdkInternalException for replay control flow. The wrapper then does:

except SdkInternalBaseException as e:
    raise Exception("Internal error during model call") from e

This conflicts with Restate’s Python error-handling guidance (re-raise internal exceptions; don’t catch them as generic Exception). Integrators must walk __cause__ chains to recover semantics.

For LLM provider non-retryable 4xx (e.g. invalid tool schema → HTTP 400), the failure should be TerminalError inside the journaled step, not a retryable transient. Otherwise Restate retries permanently invalid requests—contrary to the AI integration guide (“handle terminal errors”, bound LLM retries).

Proposed behavior:

except SdkInternalBaseException as e:
    if isinstance(e, SuspendedException):
        raise
    # Optional: inspect e.__cause__ for provider HTTP errors
    if is_non_retryable_provider_4xx(e.__cause__):
        raise TerminalError("Model request rejected by provider", status_code=400) from e.__cause__
    raise  # re-raise SdkInternalBaseException unchanged — do NOT wrap in Exception(...)

Policy suggestion for upstream defaults:

Provider status	Treatment
400, 401, 403, 404, 422, …	TerminalError (non-retryable)
408, 429	Retryable (existing transient path)
5xx	Retryable per RunOptions

Proposals:

• Stop wrapping SdkInternalBaseException in generic Exception.
• Either adopt a default provider-status policy or expose a hook on RestateAgent / RestateModelWrapper for classifying model failures before journaling.
• Document recommended RunOptions(max_attempts=…) for LLM journal steps.