Transformer surrogates compress an industrial mesh into a small set of tokens to predict 3D flow fields. Each prediction point first retrieves information from the compressed representation and then transforms the retrieved features with a feed‑forward layer. Existing backbones treat both steps as location‑agnostic. This work proposes two wall‑related conditioning mechanisms.\ \ Distance‑aware cross‑attention (DA‑CA) reshapes each volume query by its distance to the wall before retrieval, allowing points deep in the boundary layer to gather different geometric cues than points in the outer flow.\ \ Surface‑volume mixture‑of‑experts (SVMoE) replaces the shared feed‑forward layer with a small set of experts; volume points are routed by wall distance, while surface points are routed by local geometry. Both mechanisms are architecture‑agnostic and are applied unchanged to AB‑UPT and Transolver‑3.\ \ On the DrivAerML dataset with 50 training cases, DA‑CA reduces volume pressure error by 10.1%, and the combination of DA‑CA and SVMoE cuts it by 12.5%. DA‑CA improves the near‑wall region at a slight cost to the far field, which SVMoE recovers; the volume experts naturally specialize into near‑wall, transition, and free‑stream bands without explicit routing supervision.\ \ When retrained on 300 cases, conditioning improves every field quantity: AB‑UPT sees a 33.1% drop in volume pressure error and an 18.6% drop in velocity error; Transolver‑3 achieves 21.4% and 21.3% reductions respectively. In a Leave‑One‑Body‑Out evaluation on DrivAerNet++, the method lowers unseen‑body volume pressure error by up to 14.2%.\ \ Review: By explicitly modulating attention and feed‑forward layers with wall‑distance cues, the paper markedly enhances Transformer accuracy in complex 3D flow prediction, especially near the wall, highlighting the power of embedding physical priors into deep models.