Rendering untrusted text into Markdown reports safely: fences longer than any backtick run, escape prose, drop control chars
When generating a Markdown report from untrusted text (user documents, scraped evidence, LLM output), the viewer may render raw HTML and autolink URLs, so injected <script>, <img onerror>, fake # headings, links, images or pipe tables can change the report or run code.
Pattern that held up when rendered with markdown-it-py, raw HTML and linkify both on:
- **Quoted blocks (excerpts, notes):** put each in a fenced code block whose fence is one backtick longer than the longest backtick run in the text, so the text can never close it.
- **Inline identifiers (IDs, filenames, locators):** use code spans built the same way, and fold newlines so they can't start a new block.
- **Prose (answers, labels, descriptions):** backslash-escape Markdown and HTML specials line by line: \ * _ [ ] < > # + ! | ~ & plus a leading -, =, > or 1.. Also break :// and www.` so bare URLs don't autolink.
- **Control characters:** strip them.
- **Testing:** test with hostile inputs (<script>, <img onerror=...>, headings, [link](...), , |a|b| tables, long backtick runs). Assert the rendered HTML contains no tags, links, images or tables you didn't emit.
Keep rendering a pure function, with the generated-at time injectable, so output is byte-deterministic for tests.