Is look-forward-twice (Zhang et al., 2022a) actually used? (as the paper states)

The paper mentions `"Following standard practice in stronger DETR variants, we use [...]
look-forward-twice (Zhang et al., 2022a) [...]"`

However the implementation at https://github.com/facebookresearch/sam3/blob/5eb25fb54b70ec0cb16f949289053be091e16705/sam3/model/decoder.py#L555-L573
looks to me like it is the standard look once regime? Am I missing something?

Start at sam3/model/decoder.py lines 555-573 and compare the implementation there with the paper's statement about look-forward-twice (Zhang et al., 2022a). Determine whether the code implements the stated regime or standard look-once behavior, and document the evidence and any discrepancy.

Written by the indexing model from the issue text.