fix(chatgpt): use data-turn to detect upload previews vs generated images

[jesuisjohan]

Bug

opencli chatgpt image <prompt> --image a.png,b.png,c.png (2+ reference
images) can silently return the just-uploaded reference thumbnails instead
of the actual generated/edited image. With a single --image it works
correctly.

Root cause

isUserUploadPreview() in clis/chatgpt/utils.js (used by
getChatGPTVisibleImageUrls/waitForChatGPTImages to exclude a user's own
attachment previews from the "new image appeared" diff) relied on two
signals that no longer match ChatGPT's current DOM:

• turn.querySelector('h4')?.innerText — the turn heading ("You said:" /
  "ChatGPT said:") is visually hidden, so real Chrome's innerText
  (layout-dependent) resolves to '', even though .textContent correctly
  reads the heading. jsdom's innerText is always undefined, so the
  existing test suite never caught this either — it happened to keep
  passing via the aria-label/alt fallback below.
• button[aria-label^="Open image:"] — ChatGPT's current label for a
  multi-file attachment reads "Open image N of M: <name>", which no
  longer starts with "Open image:", so the selector stopped matching.

With both signals dead, classification falls through to alt-text sniffing.
Right after upload, an attachment thumbnail's alt/aria-label haven't
populated yet, so for a poll or two every uploaded image is misclassified
as "new". waitForChatGPTImages returns as soon as two consecutive polls
agree on a URL set — long enough for that transient window to win once
multiple attachments are involved, so it returns the uploads instead of
the real result.

Fix

Check the turn <section>'s own data-turn="user"|"assistant" attribute
first. It's set structurally as soon as the turn mounts — not tied to the
attachment's async metadata — so it isn't subject to the race. The
heading/aria-label checks stay as a fallback (now using textContent and
a substring aria-label match) for markup that lacks data-turn.

Verification

• Confirmed the real DOM shape live against chatgpt.com (data-turn on
  the turn section, empty innerText on the heading, "Open image N of M"
  aria-label) via opencli browser <session> eval.
• Reproduced the bug live: 3 attached reference images + a real generation
  prompt returned the 3 uploaded thumbnails as "generated" output.
• Applied this fix to the installed build and re-ran the identical
  request: it now returns exactly one file, the real generated/composed
  image.
• Added two regression tests reproducing the exact failure (multi-image
  upload racing before alt/aria-label populate, and the "Open image N of
  M" aria-label format) — both fail against the pre-fix code and pass
  after the fix.
• npx vitest run clis/chatgpt/utils.test.js clis/chatgpt/image.test.js →
  128/128 passing.

🤖 Generated with Claude Code