Problem

MetricRunner.run in src/alpamayo_r1/metrics/metric_runner.py declares -> dict[str, Any] but the body has no return statement:

def run(
    self, model: Any, data_batch: dict[str, Any], output_batch: dict[str, Any]
) -> dict[str, Any]:
    """Run the metrics one-by-one by the order of the metrics list."""
    per_sample_metrics = {}
    for metric in self.metrics:
        per_sample = metric.evaluate(model, data_batch, output_batch)
        per_sample_metrics.update(per_sample)

    output_batch.update({"metric/" + k: v for k, v in per_sample_metrics.items()})

Python returns None implicitly while the annotation lies. This trips up type checkers (mypy, pyright) and is misleading to callers who reasonably assume they should read a result from the return value.

The actual contract — documented in the class docstring ("…added to the output batch") and matching every existing call site — is to merge the prefixed metrics into output_batch in-place.

Fix

• Change the return annotation from dict[str, Any] to None so the type matches the runtime behavior.
• Expand the one-line docstring to spell out the in-place mutation contract and explicitly tell callers to read results from output_batch after the call.

No runtime behavior change.

Verification

grep -A 12 "def run" src/alpamayo_r1/metrics/metric_runner.py shows the body has no return statement; the only side effect is the output_batch.update(...) call on the last line.