Memes convey implicit pragmatic meanings through tightly coupled visual and textual cues, making their interpretation more challenging than that of conventional image-text pairs. Although large vision-language models (LVLMs) perform well on multimodal benchmarks, it remains unclear whether they infer non-literal mechanisms underlying memes, such as irony, wordplay, or rely mainly on superficial visual-textual correlations. To study this problem, we introduce MemeScope, a fine-grained benchmark of 1,781 memes annotated with eight pragmatic mechanisms and stage-wise explanations covering visual grounding, cross-modal alignment, and pragmatic inference. Building on MemeScope, we propose a unified interpretability framework that combines layer-wise probing with hidden-state patching to examine both the emergence and the causal contributions of meme-level pragmatic representations within LVLMs. Our analysis shows that literal visual and textual cues are often captured in earlier layers, whereas pragmatic mechanism information emerges more strongly in higher layers and shows stronger patching effects in late-layer multimodal representations. We further identify a recurring pragmatic drift, where models ground meme content correctly but move toward plausible yet incorrect mechanisms during deeper decoding.