fix(xiaohongshu): scan composer media from document.body so native text-to-image works

[liuxinyea]

Description

Problem

opencli xiaohongshu publish --card-text (小红书原生的「文字生成图片」/ text-to-image card flow) always aborts with:

expected at least N visible media item(s), got 0

Root cause

currentComposerMediaCount() located the composer's DOM root by walking up from the title element:

const root = titleEl?.closest('form, [class*="publish"], [class*="editor"], [class*="note"]') || document.body;

Xiaohongshu's current React renderer mounts the image / card editor in a different subtree than the title. So closest(...) resolves to a container that never contains the generated card media — the scan always returns 0, and the publisher refuses to submit even though the cards were generated correctly.

Fix

• Scan from document.body instead of the narrow closest(...) match, so generated cards are always found regardless of where XHS mounts the editor.
• Extend the media selector with image, svg (alongside img, video, canvas) for completeness.

       const titleEl = Array.from(document.querySelectorAll('input, textarea, [contenteditable="true"]'))
         .map((sel) => Array.from(document.querySelectorAll(sel)))
         .flat()
         .find((el) => visibleBox(el));
-      const root = titleEl?.closest('form, [class*="publish"], [class*="editor"], [class*="note"]') || document.body;
+      const root = document.body;
       const seen = new Set();
       let count = 0;
-      for (const el of Array.from(root.querySelectorAll('img, video, canvas, [style*="background-image"]'))) {
+      for (const el of Array.from(root.querySelectorAll('img, image, svg, video, canvas, [style*="background-image"]'))) {
         if (!visibleMedia(el)) continue;

Verification

• Manual: opencli xiaohongshu publish --card-text "..." --card-style 简约 --draft now generates the cards and passes the media-count gate (previously got 0).
• Confirmed on two independent browser profiles / accounts.

Trade-offs

• document.body is broader than the old root, but safe: the function already filters each candidate via visibleMedia(el) + getBoundingClientRect(), so unrelated images elsewhere on the page are excluded.
• If XHS later wraps the editor in a reliably-located, specific container, we can narrow root back down.

Related issue: (none — discovered via a real automation run; link one here if it exists)

Type of Change

• 🐛 Bug fix
• ✨ New feature
• 🌐 New site adapter
• 📝 Documentation
• ♻️ Refactor
• 🔧 CI / build / tooling

Checklist

• I ran the checks relevant to this PR — built the project (dist/ compiled) and verified the --card-text flow manually end-to-end.
• I updated tests or docs if needed — N/A: no unit/e2e test currently covers currentComposerMediaCount; DOM detection needs a live XHS session. (Happy to add a regression test if a headless harness is available.)
• I included output or screenshots when useful — error message + manual verification captured above.

Documentation (if adding/modifying an adapter)

• Not applicable — this is a bug fix in an existing adapter. No new command, no discoverability change, no new doc page, so the adapter-doc checklist items are intentionally left unchecked.

Screenshots / Output

Before

expected at least 1 visible media item(s), got 0

After (draft publish with --card-text)

• Cards generated by XHS native text-to-image
• Publisher proceeds past the media-count gate and creates the draft successfully