FIX Conversation drawer races

[Roman Lutz (romanlutz)]

Description

The recurring merge-queue image-recovery timeout happens before recovery, while selecting another conversation (failure log). A delayed send refresh can auto-open the drawer after the helper checks visibility. Its pending toggle click then closes the drawer, moving and detaching the conversation row.

Apply the initial drawer default once per attack so later refreshes cannot override the operator's open/closed choice. The selection helper checks aria-expanded, uses normal accessible clicks, and verifies the destination URL and selected row. Historical attacks still auto-open on desktop after asynchronous metadata loading; narrow screens remain closed until requested and close after selection. Creating, copying, and repeating conversations retain their existing drawer behavior.

No API, image-conversion, recovered-draft, or FileReader serialization changes. No forced clicks, fixed sleeps, disabled motion, timeout increases, or weakened recovery assertions.

Tests and Documentation

Adds controlled real-backend coverage for delayed send refreshes, closing drawers, and delayed historical metadata on desktop and narrow viewports. The pre-fix closing-drawer regressions reproduced the original unstable-element/detached-row timeout. Updates the frontend README with drawer defaults and selection synchronization guidance.

Validation ran from frontend on Windows. Seeded runs used the real in-memory backend with a scrubbed environment and only loopback provider fixtures with placeholder keys. No real provider APIs were called. CI merge-group runs were not rerun locally.

Command	Result
npm test -- --runTestsByPath src\components\Chat\ChatWindow.test.tsx src\components\Chat\ConversationPanel.test.tsx --runInBand	219 passed
npm run test:e2e -- --project seeded --grep "delayed send refresh|drawer is closing|preserves an image converter|repeats and nests|waits for the selected conversation" --repeat-each=3 --retries=0 --workers=1 --reporter=list --trace=on --output="C:\Users\romanlutz\.copilot\session-state\882ee962-a436-4346-ba02-d175591b3053\files\selection-final"	18 passed, including 3 repetitions of the unchanged image-recovery assertions
npm run test:e2e -- --project seeded --grep "historical drawer defaults" --repeat-each=3 --retries=0 --workers=1 --reporter=list --trace=on --output="C:\Users\romanlutz\.copilot\session-state\882ee962-a436-4346-ba02-d175591b3053\files\selection-historical"	6 passed
npm run type-check	Passed
npx eslint src\components\Chat\ChatWindow.tsx src\components\Chat\ChatWindow.test.tsx e2e\chat-recovery.spec.ts --max-warnings 0	Passed

Playwright used E2E_FRONTEND_PORT=19388, PYRIT_BACKEND_URL=http://127.0.0.1:19387, the worktree's virtual-environment Python, and CI unset. An additional E2E TypeScript comparison found exactly the same two upstream TS2339 errors for original_prompt_id in untouched repeat-history assertions, with no new diagnostics. Production type-check and commit hooks passed.

JupyText was not run because no Python documentation or notebooks changed.