Problem

The docstrings in src/alpamayo_r1/metrics/metric_utils.py and src/alpamayo_r1/metrics/distance_metrics.py advertise return-dict keys that are never produced. Anyone iterating these dicts looking for the documented _sq keys gets a KeyError.

1. summarize_metric (metric_utils.py)

Docstring claims it returns:

name: shape [B] # average
name_sq: shape [B] # average square
name_std: shape [B] # standard deviation over N

But the body only adds name_std:

if val.shape[1] > 1 and not disable_summary:
    mean_sq = val.pow(2).mean(1)
    std = (mean_sq - mean**2).sqrt()
    result[metric_name + "_std"] = std

mean_sq is computed as an intermediate of std and discarded. Also, the gating on N > 1 is not obvious from the docstring.

2. compute_minade (distance_metrics.py)

Docstring lists min_ade_sq, but the body returns min_ade plus per-horizon keys of the form min_ade/by_t={t * time_step:.1f} (then optionally _std variants via summarize_metric). The per-horizon keys are the function's headline output and were entirely undocumented.

3. compute_grouped_corner_distance (distance_metrics.py)

Docstring lists corner_distance_sq, but the body returns corner_distance (and corner_distance_std from summarize_metric when N > 1 and disable_summary=False).

Fix

Pure docstring change — no code logic touched. Aligns each docstring with the actual returned keys and documents how summarize_metric layers _std on top.

Verification

Skimmed the call sites (none rely on the documented _sq keys), and the new wording matches the runtime behavior verifiable by reading the function bodies.