[privatenumber]

Summary

• Exposes macOS 26 structured document recognition through ocrDocument() and ocrDocument.pages()
• Reuses the shared native service for main-thread document calls instead of spawning a new process per request
• Preserves worker-thread one-shot behavior and the strict document result contract

Problem

The document core and CLI can recognize paragraphs, tables, and lists, but Node callers still need a typed API. A one-shot wrapper would undo the shared-service behavior that protects server workloads from spawning a native process for every concurrent request.

For example, a server can receive ordinary OCR, document-page streaming, and searchable-PDF calls together. Each must wait in one FIFO queue so Vision work remains serial, cancellation is consistent, and the process is reused.

Changes

Add document and document-pages service operations with the existing framed unary and pull-controlled stream protocol. Node validates the recursive document payload before exposing it, including malformed candidate arrays, page metadata, and the macOS-unavailable error.

Main-thread document calls now share the native service with the existing APIs; worker calls retain their existing one-shot lifecycle. The public result schema and document language rules match the CLI.