fix(chatgpt-review): remove duplicate SELECTORS.responseActions key breaking clipboard copy

[BorisTyshkevich]

Summary

Found live during the same /ship 639 dry-run smoke test as PR #664, after
that fix's diagnostics made the real captured response text visible for the
first time.

SELECTORS (skills/chatgpt-review/scripts/lib/browser.mjs) defined
responseActions twice: the properly-scoped, commented
['[role="group"][aria-label="Response actions" i]'] (added alongside
copyLatestAssistantMarkdown), immediately shadowed a few lines later by a
pre-existing, unrelated ['button[aria-label*="Good response"]', ...]
definition that was never removed when the new one was added. JS object
literals silently let the second key win with no error and no lint warning
(this repo has no ESLint config anywhere), so copyLatestAssistantMarkdown
has been querying the wrong selector since the day it was introduced,
always returning null, and every caller silently fell back to plain
.innerText() — which strips Markdown heading syntax.

This is the real root cause of the recurring
"invalid_response: The delimited plan is not complete Markdown with a heading" failures in plan-author mode across this project's history,
including 3 reproductions in a single live dry run today. Confirmed live
against a real open ChatGPT tab (chrome-devtools MCP):

• the correct selector matched 3 [role="group"] elements, the last one
  containing exactly 1 real copy button;
• the shadowing selector's own text ("Good response"/"Bad response")
  matched zero elements in ChatGPT's current UI — doubly dead, wrong
  and stale.

Fix

Delete the shadowing duplicate key. Added two regression tests:

• SELECTORS.responseActions equals the scoped group selector directly
  (not the stale button selector) — independent of mechanics.
• A static parse of the SELECTORS object literal's source text asserting
  no top-level key repeats — can't become vacuously self-consistent the way
  the existing mock-based tests (which read SELECTORS.responseActions[0]
  to build their own mocks) are.

Test plan

• node --test skills/chatgpt-review/tests/*.test.mjs — 64/64 pass (62 + 2 new)
• Root cause confirmed live against a real ChatGPT tab via chrome-devtools MCP

🤖 Generated with Claude Code

https://claude.ai/code/session_01LwFPT465eDJqYcRa8HGNLz