摘要: Cross-species antimicrobial resistance (AMR) prediction is fundamentally an out-of-distribution generalization problem: models trained on one set of bacterial taxa must transfer to phylogenetically distinct genomes that may rely on different resistance mechanisms. Critically, resistance is not monolithic. Across species, it arises from a heterogeneous mixture of localized, horizontally transferred gene cassettes and diffuse, species-specific genomic backgrounds, making successful transfer inherently mechanism-dependent. Using a strict species holdout protocol, we first establish an interpretable k-mer baseline with Kover, showing that strong within-species performance collapses under true cross-species evaluation. This motivates the need for representation-level choices that explicitly preserve transferable biological signals rather than amplify phylogenetic shortcuts. We introduce two ingredients that make genomic foundation model embeddings effective for cross-species AMR prediction. First, for layer selection, we develop diagnostics for activation scale, isotropy, effective rank, and cross-seed stability under native bfloat16 inference. These reveal a sharp stability boundary at Layer 11 in Evo-1-8k-base, identifying Layer 10 as the deepest jointly stable layer; extracting embeddings here improves downstream conditioning, reproducibility, and robustness. Second, for feature aggregation, we argue that global pooling obscures localized resistance mechanisms. Instead, we treat per-window embeddings as an ordered multivariate signal and apply MiniRocket to summarize multi-scale local activation patterns. This preserves cassette-scale signals (e.g., plasmid-borne β-lactamases) that global averages dilute, reorganizing feature space toward phenotype-aligned neighborhoods where simple classifiers can generalize across species. On ampicillin resistance across 3,388 genomes from 126 species, we show that cross-species performance depends on which resistance mechanisms dominate the held-out species, not on aggregation method alone. MiniRocket excels when cassette-mediated resistance predominates; Global Pooling remains competitive for chromosomal or diffuse mechanisms. Both approaches perform similarly under same-species evaluation. Beyond accuracy, MiniRocket enables zero-training aggregation, interpretable predictions via neighbor auditing, and biological validation through mechanism-based clustering. Unlike complex decision boundaries learned by gradient boosting, k-NN exposes the underlying geometric reorganization that explains when and why local pattern preservation succeeds: reduced phylogenetic hubness and increased cross-species mechanism sharing. Together, our results establish aggregation choice as a central axis in cross-species AMR prediction and provide a reproducible, diagnostic-driven framework for deploying genomic foundation models under distribution shift.