mikiane/claude-watermark-cleaner ? reverse-engineered prompt

Build me a small Python tool that takes a text file and tries to clean up Claude watermark traces in a practical way.

I want it to first remove any suspicious invisible or directional Unicode characters, then lightly rewrite the text using a model that is not Claude, ideally a local Ollama model first, with Codex as a fallback if needed. It should be careful with real content, so keep code blocks, inline code, URLs, variables, and numbers unchanged. If a number changes, reject that output and try again.

It should work from the command line on a file, and also have a clipboard mode for macOS. I’d like an option to do only the cleaning step without rewriting, and another option to limit how many rewrite passes it tries. Please add tests too, using only the Python standard library, and make the script print a short summary of what engine it used, how many invisible characters were removed, and how much the text changed.