Problem

Two docstring drift bugs in DistanceMetrics (src/alpamayo_r1/metrics/metric_api.py) — same pattern that #86 already fixed for the lower-level functions in distance_metrics.py / metric_utils.py.

1. Phantom group_by_scenarios parameter on __init__

def __init__(self, prefix: str = "", time_step: float = 0.1):
    """Args:
    prefix (str, optional): ...
    group_by_scenarios (list[str], optional): Defaults to None.
        if not None, then the metrics will be grouped by the scenario names in this list.
    """

The signature has only prefix and time_step. group_by_scenarios is documented but does not exist — likely a leftover from an earlier API. Worse, time_step (which is the actual second parameter and matters for the per-horizon keys) is undocumented.

2. _sq keys in evaluate() Returns block that are never returned

Returns dict[str,Tensor]:
    min_ade: [B] average min_ade ...
    corner_distance: [B] ...
    # if N > 1, we have the following extra data item:
    # for each value above, we collect the statistics _sq and _std, for example:
    min_ade_sq: [B] average min_ade^2 ...
    min_ade_std: [B] std of min_ade

The body only emits min_ade, min_ade/by_t={H:.1f}, ade, ade/by_t=3.0, corner_distance, plus _std variants from summarize_metric when num_traj_sets > 1. The _sq keys exist nowhere; the ade* and by_t* keys (which are the bulk of the actual output) are completely undocumented.

Fix

Pure docstring change. No code logic touched.

• Replace the dead group_by_scenarios entry with an accurate description of time_step.
• Rewrite the Returns block to list every key the body actually emits, mirror the format used in Sync metric docstrings with the keys actually returned #86 for compute_minade, and call out that all keys are prefixed via apply_prefix(self.prefix, ...).

Verification

The new key list matches the body's metric_dict.update(...) calls (around lines 232 for ade, 241 for ade/by_t=3.0, and the return of compute_grouped_corner_distance for corner_distance).