Keep listening while an utterance is being transcribed

stt.finish() used to run inline in the frame loop, deafening the
assistant for up to ~4s per utterance while the server returned the
transcript (and whatever was said meanwhile was later drained). The
finish now runs on a background thread per utterance; the segmenter
keeps consuming mic frames, so speech during the wait becomes its own
utterance and is transcribed too. Transcripts are delivered oldest-
first between frames, deferred while an utterance is in progress, and
the command/follow-up window no longer times out on transcription
latency. Speaking stays half-duplex (echo protection).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

tests/test_assistant.py

@@ -111,19 +111,22 @@ def test_long_speech_is_cut_off_at_the_cap():
 # ── the flow ────────────────────────────────────────────────────────────────
 
 class FakeTranscription:
-    """Each utterance returns the next scripted transcript."""
+    """Each utterance takes the next scripted transcript. Bound at creation
+    (the "start" event, on the main thread) because finish() now runs on a
+    worker thread per utterance and two may be in flight at once."""
 
     script = []
 
     def __init__(self):
         self.frames = 0
         self.cancelled = False
+        self.transcript = FakeTranscription.script.pop(0)
 
     def add(self, frames):
         self.frames += len(frames)
 
     def finish(self, audio):
-        return FakeTranscription.script.pop(0)
+        return self.transcript
 
     def cancel(self):
         self.cancelled = True
@@ -141,7 +144,7 @@ def utterance():
     return [frame(LOUD)] * 8 + [frame(QUIET)] * 12
 
 
-def make(replies, commands="read", streaming=False, **kw):
+def make(replies, commands="read", streaming=False, transcription=FakeTranscription, **kw):
     asked, played, logs = [], [], []
 
     def converse(message, on_command, on_say=None):
@@ -152,7 +155,7 @@ def make(replies, commands="read", streaming=False, **kw):
         return server.Reply(reply)
 
     clock = Clock()
-    a = Assistant(transcription=FakeTranscription, converse=converse,
+    a = Assistant(transcription=transcription, converse=converse,
                   converse_stream=None if not streaming else kw.pop("stream"),
                   tts=lambda text, voice: [text.encode()], play=lambda chunks: played.extend(chunks),
                   tools=Tools(mode=commands, log=logs.append), log=logs.append, clock=clock, **kw)
@@ -163,6 +166,7 @@ def say(a, transcript):
     FakeTranscription.script.append(transcript)
     for f in utterance():
         a.process(f)
+    a.handle_transcripts(wait=True)  # what run() does between frames
 
 
 def spoken(played):
@@ -225,6 +229,27 @@ def test_mic_audio_during_speech_is_discarded():
     assert not a.segmenter.in_speech
 
 
+def test_listening_continues_while_a_transcript_is_in_flight():
+    import threading
+
+    release = threading.Event()
+
+    class Slow(FakeTranscription):
+        def finish(self, audio):
+            release.wait(timeout=5)
+            return self.transcript
+
+    a, asked, played, logs, _ = make(["One.", "Two."], transcription=Slow)
+    FakeTranscription.script += ["hey bolt first thing", "hey bolt second thing"]
+    for f in utterance() + utterance():  # the second said while the first transcribes
+        a.process(f)
+    assert asked == []  # both transcripts still at the "server"
+    release.set()
+    a.handle_transcripts(wait=True)
+    assert len(asked) == 2  # nothing said during the wait was lost
+    assert asked[0].startswith("first thing") and asked[1].startswith("second thing")
+
+
 def test_streamed_sentences_are_spoken_as_they_arrive():
     def stream(message, on_say, on_command):
         on_say("First.")