feat(PromptInput): restore user message to textarea after provider errors

[janet-e4]

Problem

When the LLM backend returns an error response (e.g. the provider is unreachable), the user's typed message is silently lost. This happens because:

1. clearPromptInputDraft() fires in handleSubmit before the API call
2. setMessageEmit("") clears the textarea immediately after
3. The stream error arrives asynchronously — by then both the textarea and the localStorage draft are already wiped

The user sees "Could not respond to message" with an empty input box and must re-type their entire message to retry.

Solution

Added savePromptInputDraft(storageKey, value) to usePromptInputStorage — a counterpart to the existing clearPromptInputDraft. In ChatContainer, the chatHandler passed to multiplexStream now tracks whether the stream ended with type: "abort". If it did, after multiplexStream resolves:

• setMessageEmit(promptMessage.userMessage) restores the text to the textarea
• savePromptInputDraft(storageKey, message) writes it back to localStorage

The user's message reappears in the input ready to retry — no re-typing needed.

Changes

• frontend/src/hooks/usePromptInputStorage.js — add exported savePromptInputDraft(storageKey, value) (mirrors existing clearPromptInputDraft, 15 lines)
• frontend/src/components/WorkspaceChat/ChatContainer/index.jsx — import savePromptInputDraft; wrap chatHandler in fetchReply to detect type: "abort"; restore message after stream resolves on error (14 lines)

No new dependencies. No changes to existing autosave/restore behavior (page-refresh drafts, debounced writes, thread/workspace scoping).

Test steps

1. Open any workspace in chat mode
2. Type a message
3. Temporarily break the LLM provider (e.g. set an invalid API key or base URL in Settings)
4. Submit the message
5. ✅ "Could not respond to message" error appears in chat history
6. ✅ The typed message reappears in the textarea
7. ✅ On page refresh, the draft is also still there
8. Restore the correct provider config
9. Resubmit — message sends successfully
10. ✅ Textarea clears, draft is gone, response appears

Affected providers

Provider-agnostic — triggers on any type: "abort" stream response regardless of which LLM backend is configured.