Summary

Continues the pattern of merged PR #50 (helper.create_message) and PR #73 (load_physical_aiavdataset): convert the remaining public-facing input validators from assert to ValueError, so they survive python -O and emit a clear exception type users can catch.

Changes

1. tokenize_history_trajectory (src/alpamayo_r1/models/base_model.py:104-105)

-    assert "ego_history_xyz" in traj_data
-    assert traj_data["ego_history_xyz"].ndim == 4, "ego_history_xyz must be 4D of [B, n_traj, T, 3]"
+    if "ego_history_xyz" not in traj_data:
+        raise ValueError(
+            f"`traj_data` is missing required key 'ego_history_xyz'; got keys {list(traj_data)}"
+        )
+    if traj_data["ego_history_xyz"].ndim != 4:
+        raise ValueError(
+            "ego_history_xyz must be 4D of [B, n_traj, T, 3], got "
+            f"shape {tuple(traj_data['ego_history_xyz'].shape)}"
+        )

The first assert had no message; the second one's was correct but assert got stripped under -O. Both errors now include the actual offending value (key set or shape) for faster diagnosis.

2. AlpamayoR1.sample_trajectories_from_data_with_vlm_rollout (src/alpamayo_r1/models/alpamayo_r1.py:157)

-        assert n_traj_group == 1, "Only one trajectory group is supported for inference."
+        if n_traj_group != 1:
+            raise ValueError(
+                f"Only one trajectory group is supported for inference, got {n_traj_group=}"
+            )

Scope

These three call sites — plus the two already addressed in #50 and #73 — are all the public-facing input validators that were using assert. I deliberately left the remaining asserts as-is because they guard internal invariants, not caller-supplied input:

If you'd prefer a sweeping conversion to ValueError everywhere, happy to do that in a follow-up — but the convention here mirrors the way #50 was framed (only the user-input boundary).

Test plan

• Both files parse: python -c \"import ast; [ast.parse(open(f).read()) for f in ['src/alpamayo_r1/models/base_model.py', 'src/alpamayo_r1/models/alpamayo_r1.py']]\"
• No semantics change: every check fires under exactly the same condition; just a different exception class with a more informative message.
• Behavior under python -O: the validation now runs (previously stripped). For valid inputs this is a no-op; for invalid inputs callers see ValueError instead of an opaque downstream AttributeError/KeyError.