Learning theory treats data as samples supplied by an abstract oracle, whereas a physical learner accesses environment only through a concrete experiment. Finite detector resolution, a limited number of quantum-state preparations, and restrictions on information processing determine which distinctions can reach the learner. Here, we formulate “physics-aware learnability” to keep the statistical task fixed while making the physical model and its resource budget explicit. Our central principle is operational: when different environments require incompatible decisions, successful learning is possible only if the allowed experiment preserves sufficient distinguishability between them. Based on this principle, we show that coarse-grained measurements can make a task impossible even when it is learnable under full access, while hypotheses constructed from measurement cells define a distinct quotient task. For a fixed finite partition, we identify an exact worst-case coverage trade-off between the readout alphabet size and the number of cells the learner may select, and provide a finite-sample learner. In the case of quantum access, the framework connects task conflict directly to quantum-state discrimination, yielding necessary distinguishability conditions and, for pure states, an explicit lower bound on the required number of copies. Fixed-resource feasibility in finite models can further be expressed through linear or semidefinite optimization. These results establish detector resolution, quantum copy number, and admissible information processing as physical resources that can determine whether learning is possible.