Advanced Persistent Threat (APT) actors are stealthy, adaptive, and persistent, so they often evade traditional intrusion detection systems. The clearest evidence appears during the intrusion phase, so host events must be tracked and labeled individually. No public Windows dataset does so. We present Prowl, a Windows kernel-operation trace dataset of 4,411,161 labeled events. Of these, 79.4% are file and registry operations that can be mapped to native system calls while retaining their arguments. 18.4% are periodic process resource samples, and 2.2% are process, thread, and image lifecycle callbacks. A breach-and-attack-simulation platform ran adversary tradecraft in an instrumented sandbox while Process Monitor recorded every event, giving 20 attack captures and five baselines. Existing host datasets label per trace, sample, time window, or red-team report, so benign events inside an attack capture inherit malicious labels. We label each event by execution-context provenance. Process instances are rebuilt from process and parent identifiers and chronological order. An event is malicious when its process instance descends from the attack process tree. This preserves attribution under identifier reuse and captures activity through signed system binaries. The release includes structural validation, harness-artifact scrubbing, and a label-conflict audit that sets a 99.84% accuracy ceiling. Grouped evaluation reaches 0.998 ROC-AUC, and leave-one-capture-out 0.987 +/- 0.025. However, process identity drives these scores. Process-name lookup alone reaches 0.9988 ROC-AUC, and withholding test process names drops logistic regression to 0.895. We therefore recommend identity-blind held-out-process evaluation as the primary protocol for determining whether a model learns attack behavior rather than process identity.