feat(i18n): persist language preference and thread through chat, memoir, TTS

- Add users.language_preference (Alembic 0018, default zh); capture at signup/SMS
  only; expose on auth and profile APIs
- Lite English prompts for chat and memoir; localized stage labels and agent
  names (Life Echo / 岁月知己)
- Tencent TTS: language-aware synthesis, ModelType=1 for 501004, English chunking
- WebSocket pipeline: emit all AGENT_RESPONSE segments when TTS cancels; INFO logs
  for tts_this_turn and TTS decisions; on-demand TTS logging
- Expo: device language on auth, i18n tiers/agent name, [SPLIT] streaming UX fixes
- Tests for migration, prompts, pipeline, router tts_this_turn, reply segments

Co-authored-by: Cursor <cursoragent@cursor.com>

api/app/agents/chat/prompts_conversation.py

@@ -9,8 +9,13 @@ from app.agents.chat.background_voice import (
     normalize_background_voice,
 )
 from app.agents.chat.occupation_context import get_occupation_chat_hint
-from app.agents.chat.output_rules import chat_output_rules
+from app.agents.chat.output_rules import (
+    chat_output_rules,
+    chat_output_rules_en,
+)
 from app.agents.chat.personas import (
+    AGENT_NAME_EN,
+    AGENT_NAME_ZH,
     get_interview_persona_tone_hint,
     normalize_interview_persona,
 )
@@ -23,7 +28,10 @@ from app.agents.chat.prompt_layers import (
     build_reply_strategy_block,
     build_style_profile_block,
 )
-from app.agents.stage_constants import STAGE_DISPLAY_ZH, STAGE_ERA_HINTS
+from app.agents.stage_constants import (
+    STAGE_ERA_HINTS,
+    stage_display_name,
+)
 from app.agents.state_schema import KnownFact, PersonaThread
 from app.core.config import settings
 
@@ -56,6 +64,36 @@ SLOT_NAME_MAP = {
     "lesson": "人生经验",
 }
 
+SLOT_NAME_MAP_EN = {
+    "place": "where you grew up",
+    "people": "important people",
+    "daily_life": "everyday life",
+    "emotion": "childhood feelings",
+    "turning_event": "memorable moments",
+    "school": "school experiences",
+    "city": "the city you studied in",
+    "motivation": "what drove you",
+    "challenge": "challenges you faced",
+    "change": "how you changed",
+    "job": "what you did at work",
+    "environment": "your work environment",
+    "decision": "important decisions",
+    "pressure": "pressure and hardship",
+    "growth": "career growth",
+    "relationship": "family relationships",
+    "conflict": "conflicts and resolutions",
+    "support": "mutual support",
+    "responsibility": "family responsibilities",
+    "value": "core values",
+    "regret": "regrets and acceptance",
+    "pride": "moments you're proud of",
+    "lesson": "life lessons",
+}
+
+
+def slot_name_map_for(language: str) -> Dict[str, str]:
+    return SLOT_NAME_MAP_EN if language == "en" else SLOT_NAME_MAP
+
 
 def _compact_era_hint(
     current_stage: str,
@@ -106,6 +144,61 @@ def _compact_era_hint(
     return "\n".join(parts) + "\n"
 
 
+def _get_opening_prompt_en(
+    current_stage: str,
+    empty_slots_readable: List[str],
+    user_profile_context: str = "",
+    profile_birth_year: Optional[int] = None,
+    profile_era_place: str = "",
+) -> str:
+    """English-lite opening prompt; ignores persona/background-voice nuances."""
+    stage_name = stage_display_name(current_stage, language="en")
+    if empty_slots_readable:
+        topics_str = ", ".join(empty_slots_readable)
+        topics_heading = (
+            f"## Suggested topics for this stage ({stage_name})\n"
+            f"Pick one of these to ask about: {topics_str}"
+        )
+        task_question = (
+            "2. You are a **warm, host-style confidant**: ask one **specific, "
+            "easy-to-answer, vivid** question that pulls the user into telling a "
+            "life memory; ideally land on one of the topics above. Avoid vague "
+            "openers like \"How have you been?\" Open the door with one small "
+            "anchor (a place, a person, an object, or a tiny scene from a day)."
+        )
+    else:
+        topics_heading = (
+            f"## Current stage ({stage_name})\n"
+            "The main topics for this stage are largely covered. Open with "
+            "something tied to a previous memory or a fresh small angle of this "
+            "stage; do not interrogate from the start."
+        )
+        task_question = (
+            "2. **Greeting + a memory hook**: after a warm acknowledgement, "
+            "drop a light, concrete question tied to recollection — never "
+            "small-talk filler."
+        )
+
+    profile_section = ""
+    if user_profile_context.strip():
+        profile_section = "## About the user\n" + user_profile_context.strip() + "\n"
+
+    return f"""You are "{AGENT_NAME_EN}" — a warm host-style friend. The user just opened the chat and **has not said anything yet**; you speak first. Tone like an old friend, but your job is to help the user start telling their life story; in two or three short sentences, give a greeting plus **one vivid, recollection-oriented question** tied to the current stage or suggested topics. No flowery prose, no long literary descriptions, no generic small-talk.
+
+{profile_section}{topics_heading}
+
+## Task
+1. Brief greeting.
+{task_question}
+3. Sound natural and warm.
+
+## Format
+- Use `[SPLIT]` to break into at most two short bubbles, or keep greeting + question in one short bubble.
+- {chat_output_rules_en()} Do not write the user's answer for them.
+
+Output (spoken-style English only, no Markdown):"""
+
+
 def get_opening_prompt(
     current_stage: str,
     empty_slots_readable: List[str],
@@ -115,9 +208,18 @@ def get_opening_prompt(
     occupation: str = "",
     profile_birth_year: Optional[int] = None,
     profile_era_place: str = "",
+    language: str = "zh",
 ) -> str:
     """空对话时 AI 先开口的提示词"""
-    stage_name = STAGE_DISPLAY_ZH.get(current_stage, current_stage)
+    if language == "en":
+        return _get_opening_prompt_en(
+            current_stage,
+            empty_slots_readable,
+            user_profile_context=user_profile_context,
+            profile_birth_year=profile_birth_year,
+            profile_era_place=profile_era_place,
+        )
+    stage_name = stage_display_name(current_stage, language="zh")
     bv_open = normalize_background_voice(background_voice)
     if empty_slots_readable:
         topics_str = "、".join(empty_slots_readable)
@@ -178,13 +280,13 @@ def get_opening_prompt(
         tone_paragraph = " " + " ".join(tone_bits) + "\n\n"
 
     opening_head = (
-        "你是「岁月知己」——主持式知己：用户刚进对话，**还没说话**，请你先开口。"
+        f"你是「{AGENT_NAME_ZH}」——主持式知己：用户刚进对话，**还没说话**，请你先开口。"
         "语气像老朋友，但**职责是帮对方开口讲人生故事**；两三句内问候 + **一个落在当前阶段或建议话题上的、有画面感的问题**；"
         "不要排比、不要长段文学描写，**不要**把泛泛问近况当主菜。\n\n"
     )
     if bv_open != "default":
         opening_head = (
-            "你是「岁月知己」——主持式知己：用户刚进对话，**还没说话**，请你先开口。"
+            f"你是「{AGENT_NAME_ZH}」——主持式知己：用户刚进对话，**还没说话**，请你先开口。"
             "**短**；两三句内问候 + **一个回忆向的具体问题**；不要排比、不要文学描写。\n\n"
         )
 
@@ -217,6 +319,92 @@ def get_opening_prompt(
 直接输出（仅自然口语，无 Markdown）："""
 
 
+def _get_guided_conversation_prompt_en(
+    current_stage: str,
+    empty_slots: List[str],
+    filled_slots: Dict[str, str],
+    detected_user_stage: str = "",
+    user_profile_context: str = "",
+    memory_evidence_text: str = "",
+    recent_questions: list[str] | None = None,
+    turn_directive_block: str = "",
+) -> str:
+    """English-lite guided interview prompt (no persona/voice nuances)."""
+    stage_name = stage_display_name(current_stage, language="en")
+    detected_name = (
+        stage_display_name(detected_user_stage, language="en")
+        if detected_user_stage and detected_user_stage != current_stage
+        else ""
+    )
+    empty_readable = [SLOT_NAME_MAP_EN.get(s, s) for s in empty_slots]
+    filled_lines = []
+    for k, v in (filled_slots or {}).items():
+        name = SLOT_NAME_MAP_EN.get(k, k)
+        if v:
+            filled_lines.append(f"- {name}: {v}")
+    filled_block = "\n".join(filled_lines) if filled_lines else "(none yet)"
+
+    suggested_block = (
+        "Suggested still-open angles for this stage: " + ", ".join(empty_readable)
+        if empty_readable
+        else "Main angles for this stage are largely covered."
+    )
+
+    detected_line = (
+        f"\nThe user is currently talking about: **{detected_name}** (system was tracking **{stage_name}**)."
+        if detected_name
+        else ""
+    )
+
+    profile_section = ""
+    if user_profile_context.strip():
+        profile_section = "\n## About the user\n" + user_profile_context.strip()
+
+    memory_section = ""
+    if (memory_evidence_text or "").strip():
+        memory_section = (
+            "\n## Reference memory snippets (for continuity only — do NOT write them as the user's first-person experience this turn)\n"
+            + memory_evidence_text.strip()
+        )
+
+    recent_q_section = ""
+    if recent_questions:
+        last = recent_questions[-4:]
+        recent_q_section = (
+            "\n## Recently asked questions (do NOT repeat these; offer a new angle)\n"
+            + "\n".join(f"- {q}" for q in last)
+        )
+
+    directive_block = (turn_directive_block or "").strip()
+    directive_section = (
+        f"\n## This turn's plan\n{directive_block}\n" if directive_block else ""
+    )
+
+    return f"""{directive_section}You are "{AGENT_NAME_EN}," a warm host-style friend helping the user record a memoir. Reply in conversational English.
+
+## Stage context
+Currently tracking life stage: **{stage_name}**.{detected_line}
+{suggested_block}
+
+## Already gathered for this stage
+{filled_block}{profile_section}{memory_section}{recent_q_section}
+
+## Behaviour
+- Pick up the **specific** detail the user just said (one tangible noun or short phrase) and gently push one step deeper before asking your next question.
+- Prefer ONE clear, specific question per reply. Open-ended over forced A/B options.
+- If the user is in the middle of a story, follow that thread; do not switch topics for the sake of coverage.
+- If you previously asked about something and the user already answered, do not re-ask.
+- Stay short and precise. One acknowledgement sentence + one question is the default shape.
+
+## Strict rules
+- {chat_output_rules_en()}
+
+## Format
+- Use `[SPLIT]` to split into at most two short bubbles when natural.
+
+Reply in English only. Do not output Markdown headings."""
+
+
 def get_guided_conversation_prompt(
     current_stage: str,
     empty_slots: List[str],
@@ -234,8 +422,20 @@ def get_guided_conversation_prompt(
     persona_threads: list[PersonaThread] | None = None,
     recent_questions: list[str] | None = None,
     turn_directive_block: str = "",
+    language: str = "zh",
 ) -> str:
     """生成状态感知的对话提示词；用户原话仅以 HumanMessage 传入，不写入本 system 文本。"""
+    if language == "en":
+        return _get_guided_conversation_prompt_en(
+            current_stage=current_stage,
+            empty_slots=empty_slots,
+            filled_slots=filled_slots,
+            detected_user_stage=detected_user_stage,
+            user_profile_context=user_profile_context,
+            memory_evidence_text=memory_evidence_text,
+            recent_questions=recent_questions,
+            turn_directive_block=turn_directive_block,
+        )
     persona_key = normalize_interview_persona(persona)
     persona_tone = get_interview_persona_tone_hint(persona_key)
     voice_tone = get_background_voice_tone_hint(background_voice)
@@ -307,6 +507,8 @@ def get_guided_conversation_prompt(
 
 __all__ = [
     "SLOT_NAME_MAP",
+    "SLOT_NAME_MAP_EN",
+    "slot_name_map_for",
     "get_guided_conversation_prompt",
     "get_opening_prompt",
 ]