v0.1.0

First release.

Added

• Read an image of LaTeX-rendered content and put its LaTeX on the clipboard: single formulas, sentences with inline math, numbered and unnumbered display equations, lists, headings and tables.
• A small window (a Capture button, wrap and engine options, the output text) and a texcopy command with JSON output.
• Two engines: a 118 MB formula model that runs in the program, and an optional vision model (OvisOCR2 through a local llama-server) with a guided, checksum-verified download.
• The window loads the vision engine while you drag the box, and unloads it when it's idle.
• Region selection through the OS's own tool (macOS screencapture, the common Linux screenshot tools) or texcopy's built-in overlay.
• Downloads for macOS (.app), Windows (a zip with a launcher that has no console) and Linux (a tarball with an installer), built by a release workflow.
• A benchmark that re-renders the output and compares it with the original (tools/bench.py).
• A seeded case generator written in Rust (tools/gen) that makes benchmark pictures with real LaTeX tools, and texcopy --batch to read many images with one loaded engine.
• Downloads include a license notices file, generated by tools/third_party_licenses.py.

Known issues

• The output will be wrong or badly formed in plenty of cases, so proofread it. docs/BENCHMARKS.md lists the kinds I know about, and issues and pull requests are welcome.
• I've only run it by hand on macOS (Apple Silicon). The Windows and Linux builds are built and tested by CI but haven't been tried on real machines yet.
• Styling that isn't part of the content (bold, italic, vertical table rules) doesn't come back.