@phamkhachoabk/dsh-ocr-apple-vision

Apple Vision provider for ctx.ocr. Registers two engines:

• apple-vision-documents (tier 1) — RecognizeDocumentsRequest, macOS 26+. Titles, paragraphs, lists, tables with row and column spans, barcodes, and reading order.
• apple-vision-text (tier 2) — VNRecognizeTextRequest, macOS 13+. Lines only, reported as a warning rather than passed off as structure.

Recognition runs in a prebuilt Swift sidecar, not in the harness process: a wedged Vision call is then a killable process tree rather than a blocked host, and the binary needs no Node ABI rebuilds.

Cold start

The first recognition after this binary changes blocks in _ANEClient compileModel: while the Apple Neural Engine compiles its model. Measured on macOS 27: 241s, then 98s, then ~0.3s from there on. The cost is per binary build, not per machine, so a published release pays it once per install. The plugin runs a warm-up against a tiny bundled image in the background at startup, under warmupTimeoutMs, so a user's first real image does not pay for it.

Fail-closed

A missing binary, an unsupported architecture and an OS that lacks the API all answer unusable, so consumers have one path. When no sidecar can be located the plugin logs once and registers nothing; the harness runs on without OCR.