fix(tui-v2): skip ask_user picker when candidates list is empty (closes #615)

[Kailigithub]

Summary

Issue #615: when ask_user is called with candidates=[],
_drain_ask_user_events in frontends/tuiapp_v2.py still mounted a
ChoiceList containing only the "Type something" free-text escape hatch.
That widget calls .focus() on mount, which steals keyboard focus from
the text input. The agent's empty-candidate prompt then becomes
unreachable from the keyboard.

Fix

Add a defensive early-return at the drain layer when candidates is
empty, mirroring tui_v3 commit dfab299. The 'Waiting for your answer ...' marker is already in scrollback as part of the assistant
stream, so the user replies via the normal input box without a
redirected picker.

Note: _install_ask_user_hook already filters empty candidates at the
producer side (line 3983). This drain-side guard is belt-and-suspenders
— any future producer that bypasses the hook (or a stale event pushed
before the hook was installed) will still be safe.

Diff

--- a/frontends/tuiapp_v2.py
+++ b/frontends/tuiapp_v2.py
@@ -6164,6 +6164,12 @@
             except queue.Empty: break
         if not latest: return
         question = latest["question"]; candidates = latest["candidates"]
+        # v2 parity with tui_v3 (dfab299): when the agent didn't supply
+        # candidates, mounting an empty ChoiceList still calls .focus() on
+        # mount and steals keyboard focus from the text input. Skip the
+        # card; the 'Waiting for your answer ...' marker is already in
+        # scrollback, and the user replies via the normal input box.
+        if not candidates: return
         multi = bool(self._MULTI_RE.search(question))
         kind = "multi_choice" if multi else "choice"
         choices = [(c, c) for c in candidates] + [(FREE_TEXT_LABEL, FREE_TEXT_CHOICE)]

Verification

tests/test_tuiapp_v2_ask_user_drain.py covers three cases:

1. candidates=[] → no ChatMessage appended (bug case)
2. candidates=['A','B'] → exactly one picker mounted, with the
   free-text escape hatch appended (happy-path regression)
3. candidates=[] with [多选] marker → still skipped (multi phrasing
   must not bypass the guard)

The first test fails on main (mounts the empty ChoiceList) and passes
with this fix, confirming the regression test catches the bug.

Closes #615