fix(prediction): output_iterator silently yields individual characters for string output

[devteamaegis]

Summary

output_iterator() and async_output_iterator() use self.output or [] to coerce the model's output to a list. This is only safe when output is None (not started yet) or an actual list.

For non-array model outputs — plain strings, URLs, dicts — the falsy-or pattern silently returns the raw value unchanged:

# output_iterator with output = "hello world"
previous_output = self.output or []  # → "hello world"
# ...
output = self.output or []           # → "hello world"
new_output = output[len(previous_output):]  # string slice, not list slice
yield from new_output                # silently yields 'h', 'e', 'l', 'l', ...

Fix

Replace value or [] with an explicit isinstance check:

def _as_list(value):
    if value is None:
        return []
    if isinstance(value, list):
        return value
    raise ValueError(
        f"output_iterator requires an array output type, "
        f"but the model returned a {type(value).__name__!r}. "
        f"Use prediction.output directly for non-array outputs."
    )

This resolves the long-standing # TODO: check output is list comments in both methods.

Bonus fix

The async_output_iterator loop also had a variable-shadowing bug:

# Before — outer `output` list is clobbered by the loop variable
for output in new_output:
    yield output

# After
for item in new_output:
    yield item

Tests

Six new unit tests added to tests/test_prediction.py covering:

• None output → empty iteration (no error)
• String output → ValueError (sync and async)
• Dict output → ValueError (sync)
• Completed prediction with list output → empty iteration (all tokens were "seen" at start)