feat(voice): speech-interrupted latch in the TTS streaming core

mark_speech_interrupted() / take_speech_interrupted(): a one-shot,
TTL'd (120s) latch plus SPEECH_INTERRUPTED_NOTE. Barge-in paths mark
it when they cut live speech; the next turn's submit path pops it and
prepends the note to the model-bound message — API-call local, never
persisted, so history and prompt caching are untouched.

tests/tools/test_tts_streaming.py

@@ -53,6 +53,26 @@ class TestSentenceChunker:
         ]
 
 
+# ── Interruption latch ───────────────────────────────────────────────────
+
+
+class TestSpeechInterruptedLatch:
+    def test_take_pops_and_reports_recent_barge(self):
+        ts.mark_speech_interrupted()
+        assert ts.take_speech_interrupted() is True
+        assert ts.take_speech_interrupted() is False  # one-shot
+
+    def test_untouched_latch_is_false(self):
+        ts._interrupted_at = None
+        assert ts.take_speech_interrupted() is False
+
+    def test_stale_barge_expires(self, monkeypatch):
+        ts.mark_speech_interrupted()
+        at = ts._interrupted_at
+        monkeypatch.setattr(ts.time, "monotonic", lambda: at + ts._INTERRUPT_TTL_S + 1)
+        assert ts.take_speech_interrupted() is False
+
+
 # ── Registry + resolver ──────────────────────────────────────────────────
 
 

tools/tts_streaming.py

@@ -23,6 +23,7 @@ from __future__ import annotations
 
 import logging
 import re
+import time
 from abc import ABC, abstractmethod
 from typing import Callable, Dict, Iterator, List, Optional
 
@@ -30,6 +31,34 @@ from tools.tts_tool import _get_provider, get_env_value
 
 logger = logging.getLogger(__name__)
 
+
+# ---------------------------------------------------------------------------
+# Interruption latch — lets the model know it was cut off mid-speech
+# ---------------------------------------------------------------------------
+# When the user barges in on a spoken reply (talks over it, types, hits the
+# record key), the surface marks the latch; the next turn's submit path takes
+# it and prepends SPEECH_INTERRUPTED_NOTE to the model-bound message (API-call
+# local — never persisted, same as the CLI's model-switch notes). The TTL
+# keeps a stale barge from annotating an unrelated message minutes later.
+
+SPEECH_INTERRUPTED_NOTE = (
+    "[Note: the user interrupted your previous spoken reply before it finished.]"
+)
+_INTERRUPT_TTL_S = 120.0
+_interrupted_at: Optional[float] = None
+
+
+def mark_speech_interrupted() -> None:
+    global _interrupted_at
+    _interrupted_at = time.monotonic()
+
+
+def take_speech_interrupted() -> bool:
+    """Pop the latch; True when a barge happened within the TTL."""
+    global _interrupted_at
+    at, _interrupted_at = _interrupted_at, None
+    return at is not None and time.monotonic() - at < _INTERRUPT_TTL_S
+
 # Sentence boundary: after .!? followed by whitespace, or a blank line.
 SENTENCE_BOUNDARY_RE = re.compile(r"(?<=[.!?])(?:\s|\n)|(?:\n\n)")
 _THINK_BLOCK_RE = re.compile(r"<think[\s>].*?</think>", flags=re.DOTALL)