Why

PR #86 corrected the docstrings of compute_minade, compute_grouped_corner_distance, and summarize_metric to match the keys the functions actually return. Those docstrings now act as a public contract — but there is nothing pinning that contract down. A future edit can quietly change the key names (or the gating on _std) and nothing fails.

This PR adds focused, dependency-light pytest cases that lock in the contract documented in #86.

What

src/alpamayo_r1/metrics/test_distance_metrics.py — 9 tests, no GPU / no HF auth required:

compute_ade:

• Returns the documented [B, N, K] shape.
• timestep_horizon actually truncates the time axis (head-only mean ≠ full-trajectory mean).

compute_minade:

• Returns min_ade plus one min_ade/by_t={H:.1f} per valid horizon (precision matches f"{t * time_step:.1f}").
• Returns matching _std entries when N > 1.
• Skips horizons that exceed T (the valid_horizons = [t for t in timestep_horizons if t <= T] filter).
• disable_summary=True suppresses _std.

summarize_metric:

• Adds <key>_std when N > 1 and disable_summary=False.
• Does not add _std when N == 1.
• Does not add _std when disable_summary=True, even with N > 1.
• Raises ValueError on inputs that are not [B, N].

Run

pytest src/alpamayo_r1/metrics/test_distance_metrics.py -v

Pure torch, runs on CPU in seconds. Uses torch.manual_seed(0) for determinism.