Selective On-Device File Routing over User-Defined Taxonomies

File taxonomies are project and user-specific, so a router cannot assume a fixed label set. We study selective file routing over user-defined taxonomies. Given descriptions of a file and its candidate folders at inference time, the router chooses an acceptable destination or abstains. Our compact router is a 23.0M-parameter MiniLM dual encoder with separate residual transformations for file and folder embeddings, folder prototypes, and a virtual no-match option. We fixed its operating margin on separate development workspaces, then evaluated the compact router once on T4, a sealed, held-out 200-task evaluation set from ten unseen workspaces that includes 30 no-fit and 20 multi-acceptable tasks. It reaches 89.0% action accuracy (95% workspace-cluster CI 85.0–93.0) versus 69.0% for the pretrained MiniLM baseline and reduces AURC from 0.1300 to 0.0637. At much larger scale, Qwen 3.6 27B and Gemini 3.1 Pro High reach 99.0% and 99.5%. On one 16 GB Apple M4 Mac mini, cached interactive routing averages 20.3 ms on CPU and 15.1 ms through MPS, with identical decisions across devices. A FLOAT32 Core ML deployment on a physical iPhone 17 reproduced all 200 frozen T4 decisions and averaged 11.2 ms per route, or 89.6 routes/s. The checkpoint is 92.1 MB. No-fit recall is 73.3%, so deployments should require user review before moving files.