fix: preserve confirmed followup snapshots and expand flow checks

scripts/generate_followup_audio_fixtures.py

@@ -33,19 +33,21 @@ def main():
     output.mkdir(parents=True, exist_ok=True)
     output.chmod(0o700)
     manifest = {"generator": "followup-audio-synthetic-v1", "synthetic": True,
-                "recorded_at": "2026-09-29 10:00:00", "fixtures": []}
+                "recorded_at": "2026-09-29 10:00:00", "speech_revision": "ordinary-dialogue-v2", "fixtures": []}
     words = ["紫色海豚", "水晶风车", "绿色蜻蜓", "银色树叶", "蓝色鲸鱼", "金色河流", "红色石榴", "白色云朵"]
     with tempfile.TemporaryDirectory(prefix="followup-audio-synthetic-") as temporary:
         temp = pathlib.Path(temporary)
-        middle_text = "这是纯合成的声音长度测试，不对应任何真实患者。现在继续校验这段音频的可听性。重复的测试语句不是医疗事实，不应反复生成随访记录。"
+        # Synthetic provenance stays in the manifest. The speech must not contradict
+        # the expected task by declaring all reported measurements non-facts.
+        middle_text = "客服说，电话里的声音还清楚吗？患者说，很清楚。客服说，好的，我们继续聊。患者说，好的，您说吧。"
         run(["say", "-v", args.voice, "-r", "190", "-o", str(temp / "middle.aiff"), middle_text])
         for case, seconds, systolic, diastolic in (("short", 60, 131, 86), ("medium", 900, 139, 88), ("long", 3598, 147, 91)):
             head_canary = "".join(secrets.SystemRandom().sample(words, 3))
             tail_canary = "".join(secrets.SystemRandom().sample(words, 3))
             while tail_canary == head_canary:
                 tail_canary = "".join(secrets.SystemRandom().sample(words, 3))
-            head = f"这是一段人工合成的音频测试。开始校验词是，{head_canary}。昨天晚上九点，收缩压一百二十六，舒张压八十二。前天下午两点，餐后血糖七点二。"
-            tail = f"现在是这段音频的最后部分。今天早晨七点，收缩压{systolic}，舒张压{diastolic}。末尾校验词是，{tail_canary}。合成测试到此结束。"
+            head = f"患者说，我先读三个词，{head_canary}。昨天晚上九点，我测的血压是收缩压一百二十六，舒张压八十二。前天下午两点，我测的餐后血糖是七点二。"
+            tail = f"患者说，我再补充一次测量。今天早晨七点，我的收缩压是{systolic}，舒张压是{diastolic}。最后再读三个词，{tail_canary}。客服说，好的，再见。"
             run(["say", "-v", args.voice, "-r", "190", "-o", str(temp / "head.aiff"), head])
             run(["say", "-v", args.voice, "-r", "190", "-o", str(temp / "tail.aiff"), tail])
             filler_seconds = seconds - duration(temp / "head.aiff") - duration(temp / "tail.aiff")