Static web version (docs/)

A plain HTML/JS reimplementation that talks directly to a stock vLLM server (no custom inference code). One page (docs/index.html, UI in app.js) with two modes, Highlights and Predictive text. Predictive text defaults to rewriting a document, and can continue as a multi-turn conversation where you write both sides. Two interchangeable engines:

• engine-tokens.js (default): highlights use prompt_logprobs over the document placed in the assistant turn; predictions use next-token logprobs plus a short greedy lookahead to finish words.
• engine-structured.js: asks the model, via chat completions with a JSON schema, to list edits or suggest next phrases. Not available in conversation view.

To avoid the chat template trimming whitespace, the response text never goes through the template: we render the chat prefix once (/tokenize with add_generation_prompt, then /detokenize), cache it, and send prefix + response as a text prompt. vLLM parses the special-token strings in it, giving the same token ids as tokenizing the two parts separately.

Run locally with cd docs && python3 -m http.server, then open http://localhost:8000. Add ?server=https://your-vllm-host to the URL to use a different server.