clipsmith-ocrlisted

# clipsmith-ocr Deterministic local OCR executor backed by Apple Vision.framework. No network calls, no third-party cloud, no API keys. Recognition runs on-device using the same ML model as Preview and Live Text. Two implementation details that affect output and are not in Apple's public docs: - **Bounding-box coordinate origin is lower-left** (not upper-left like screen coordinates). The script compensates by sorting on `-(y + height)` descending, then `x` ascending to produce correct reading order. - **Language correction (`setUsesLanguageCorrection_`) is always enabled** — Vision silently applies a language model to fix likely OCR errors based on the primary recognition language. This improves accuracy but can silently "correct" uncommon proper nouns or technical terms to more common alternatives. ## NEVER - **NEVER pass a remote URL or HTTP path** — Vision.framework requires a local `file://` URI. Remote paths produce empty output with no error raised. - **NEVER treat empty output as success** — Vision returns an empty string (not an exception) when: image resolution is too low, format is unsupported, or text contrast is insufficient. Always check that output is non-empty when text is expected. - **NEVER use `fast` mode for Chinese-heavy images** — `fast` drops strokes on dense CJK characters; use `accurate` (default) unless throughput is more important than accuracy. - **NEVER assume pyobjc is available system-wide** — this skill is a uv project. The execution wrappe