fix(voice): catch short echoed fragments of longer multi-sentence TTS replies

is_tts_echo() compared the captured transcript against the *entire*
spoken text with a whole-string similarity ratio, which only scores high
when the two strings are close in length. But a playback-phase barge
capture is cut immediately when the trigger fires and only spans the
pre-roll buffer plus time-to-silence, so a genuine self-capture is
typically a short fragment of a longer reply, not a near-verbatim repeat
of the whole thing -- for any response longer than a clause, the
length-diluted ratio fell below threshold and the echo sailed through
ungated.

When the whole-string check misses, also slide a window sized to the
transcript's word count across the spoken text and compare against each
window, so a short fragment echoed from within a much longer
multi-sentence reply is still caught. Add regression tests for a short
fragment matched at the start and in the middle of a longer reply.

tests/tools/test_voice_tts_echo_guard.py

@@ -39,6 +39,33 @@ class TestIsTtsEcho:
         transcript = "stop"
         assert is_tts_echo(transcript, spoken) is False
 
+    def test_short_fragment_of_longer_multi_sentence_reply_is_echo(self):
+        # Playback-phase captures are cut immediately on trigger and only
+        # span pre-roll + time-to-silence, so a real self-capture is
+        # typically a short fragment of a much longer spoken reply, not a
+        # near-verbatim repeat of the whole thing. A whole-string ratio
+        # dilutes with the length mismatch and misses this case (#75780
+        # review).
+        spoken = (
+            "Sure, here's a summary of what we found. The build failed "
+            "because of a missing dependency in the lockfile. I've already "
+            "gone ahead and regenerated it, and the tests are passing "
+            "again locally. Let me know if you'd like me to open a PR for "
+            "this or if you want to review the diff first before I do "
+            "anything else."
+        )
+        transcript = "Sure, here's a summary of what we found."
+        assert is_tts_echo(transcript, spoken) is True
+
+    def test_short_fragment_from_middle_of_reply_is_echo(self):
+        spoken = (
+            "The deployment finished successfully. All three services "
+            "came up healthy, and the smoke tests passed without any "
+            "errors."
+        )
+        transcript = "the smoke tests passed without any errors"
+        assert is_tts_echo(transcript, spoken) is True
+
     def test_empty_inputs_are_not_echo(self):
         assert is_tts_echo("", "hello") is False
         assert is_tts_echo("hello", "") is False

tools/voice_mode.py

@@ -1338,6 +1338,16 @@ def is_tts_echo(
     a strong signal of speaker-bleed self-capture (fail-closed guard for the
     playback-phase full-duplex listener, which has no acoustic echo
     cancellation; see #75780).
+
+    The playback-phase capture is cut immediately when the barge trigger
+    fires and only spans the pre-roll buffer plus time-to-silence, so for
+    any spoken reply longer than a clause the transcript is a short
+    FRAGMENT of `spoken_text`, not a near-verbatim repeat of the whole
+    thing. A whole-string ratio dilutes towards 0 as `spoken_text` grows
+    past the fragment's length, so when the whole-string check misses, we
+    also slide a window sized to the transcript's word count across
+    `spoken_text` and compare against each window, catching a short
+    fragment echoed from within a much longer multi-sentence reply.
     """
     if not transcript or not spoken_text:
         return False
@@ -1345,7 +1355,19 @@ def is_tts_echo(
     b = _normalize_for_echo_compare(spoken_text)
     if not a or not b:
         return False
-    return difflib.SequenceMatcher(None, a, b).ratio() >= threshold
+    if difflib.SequenceMatcher(None, a, b).ratio() >= threshold:
+        return True
+    b_words = b.split(" ")
+    a_word_count = len(a.split(" "))
+    if a_word_count >= len(b_words):
+        return False
+    best_ratio = 0.0
+    for start in range(0, len(b_words) - a_word_count + 1):
+        window = " ".join(b_words[start : start + a_word_count])
+        ratio = difflib.SequenceMatcher(None, a, window).ratio()
+        if ratio > best_ratio:
+            best_ratio = ratio
+    return best_ratio >= threshold
 
 
 def voice_stop_hint() -> str: