ligand_conditioning is a pure assertion, not a gate -- real ligand tensors are injected regardless of its value

[maraxen]

Finding

SamplingSpecification.ligand_conditioning does not gate anything. In _sampling_helper.py::_prepare_ligand_context (~:277-290):

if Y is None or Y_t is None or Y_m is None:
    if spec.ligand_conditioning:
        raise ValueError("ligand_conditioning=True requires ligand context tensors ...")
    Y = jnp.zeros(...)  # placeholder
    ...

ligand_conditioning is only consulted to decide whether to RAISE when ligand tensors are absent. If real tensors ARE present (via ligand_context_path, or already on the batch), they are injected into the bundle regardless of ligand_conditioning's value — ligand_conditioning=False with real tensors present still gets real ligand conditioning; there is no code path where this field actually suppresses ligand context that would otherwise be there.

Impact

A caller who explicitly sets ligand_conditioning=False expecting an unconditioned ablation, but who (directly or via a campaign row) also has ligand_context_path set or ligand tensors already on the batch, silently gets full ligand conditioning anyway. Conversely ligand_conditioning=True with no tensors present correctly raises — so the field behaves like a one-directional assertion, not a two-way switch.

Found during the 260715/260716 aminx campaign control-knob audit (same investigation that produced #110/#111/#113), while checking which declared knobs actually gate their claimed behavior.

Recommendation

Either make ligand_conditioning=False actually zero out/ignore real tensors when present (so it's a genuine ablation switch), or rename/document the field as what it currently is (a presence-assertion, not a conditioning toggle) so callers don't rely on it to suppress ligand context.