fix(slides): sanitize model-generated HTML before rendering it

[carfeii]

The Slides view builds a prompt asking the model to generate a single HTML slide, extracts HTML from the response with a few regexes, and renders the result via dangerouslySetInnerHTML with no sanitization. Repository content can steer what the model outputs here (prompt injection), and a slide whose own output already contains a <style> tag skips the page's default-styling wrapper entirely, using the model's raw output as-is. An onerror/onclick handler or <script> tag the model is induced to emit executes in the DeepWiki origin when the victim opens that repository's Slides view.

Sanitizes the extracted slide body with DOMPurify immediately after extraction (before the default-styling wrapper, whose own trusted mermaid/Chart.js init script is unaffected since it's appended after sanitization), and strips all tags from the slide title (also model-derived) before interpolating it into <h1>.

dompurify was already a transitive dependency (pulled in by mermaid for its own output); this adds it as a direct dependency at the same version range already resolved in yarn.lock, so no new package version needed locking.

Test plan

• Validated with the real dompurify package against the exact payload shape demonstrated for this finding (an img onerror handler and a script tag), confirming both are stripped.
• npx tsc --noEmit: passes.
• npx next build: full production build passes.
• No frontend test runner exists in this repo to add a persistent automated test alongside this fix.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.