Pretraining on large-scale datasets has been shown to improve transformer generalizability, even for out-of-domain (OOD) modalities and tasks. However, two common assumptions often fail under OOD transfer: that downstream datasets can be adapted to the fixed input geometry of pretrained models and that pretrained representations transfer effectively across imaging modalities. We show that these assumptions break down through two interacting failure modes in CT-to-MRI transfer: inefficient token usage caused by zero-padding to match pretrained input dimensions and ineffective feature adaptation. These failures led to accuracy degradation despite extensive fine-tuning. We investigated these failure modes using two CT-pretrained hierarchical shifted-window transformer backbones, SMIT and Swin UNETR, pretrained with different objectives and datasets. Mechanistic analysis introduced an attention dilution index (ADI), an entropy-based metric quantifying attention diverted toward uninformative padding tokens, and centered kernel alignment (CKA) to measure feature reuse in MRI tasks. ADI increased with zero-padding, while high feature reuse did not necessarily correspond to improved accuracy. To mitigate these issues, we introduced two interventions: a tumor-aware augmentation strategy to improve tumor appearance heterogeneity coverage and an anisotropic cropping strategy to restore token efficiency. Fine-tuning on identical rectal MRI datasets improved detection rates to 224/247 (90.7%) for SMIT and 219/247 (88.7%) for Swin UNETR, demonstrating improved robustness under CT-to-MRI transfer. This study is among the first to examine when pretrained transformers fail to transfer effectively across imaging modalities and how simple mitigation strategies, motivated by mechanistic analysis of datasets, can reduce transfer limitations while improving robustness and MRI detection.