✨ feat: WP-05 模型网关核心——逻辑模型、流式适配与 Run 真实执行

- 迁移 7+8：logical_model（能力目录）、provider_endpoint（密钥只存环境变量名，永不落库）、turn、model_call（每 attempt 单独持久化：usage/TTFT/错误，Provider 未返回保持 null 不估算——ADR-003）；message.run_id 关联
- OpenAI-compatible 流式适配器：httpx SSE 解析、标准化 StreamChunk（delta/usage/done）、transport 可注入（测试与故障注入）；密钥经环境变量读取不进日志
- Run 执行器：后台执行完整生命周期——running → content.block.started → 真实 content.block.delta*（流式产生，delta_index 单调）→ content.block.completed → message.completed → 终态；assistant 消息为权威快照；ModelCall 记录 usage 与 TTFT；无可用端点时 Run 显式失败并留 error 事件（fail-closed）
- post_message：响应提交后由框架 BackgroundTasks 执行 Run（显式预提交保证 Run 落库）；SSE 改为增量轮询（重放 + 等待新事件至终态，30s 上限）
- 测试 54 项全绿（+2）：端到端 mock 流验证 ModelCall usage/TTFT/assistant 权威消息、无端点 Run failed + content.block.failed 事件

backend/app/conversations/routes.py

@@ -4,7 +4,7 @@ import json
 from collections.abc import AsyncIterator
 from typing import Annotated
 
-from fastapi import APIRouter, Depends, Header, HTTPException, Request, status
+from fastapi import APIRouter, BackgroundTasks, Depends, Header, HTTPException, Request, status
 from fastapi.responses import JSONResponse, StreamingResponse
 from sqlalchemy import select
 
@@ -51,6 +51,7 @@ async def post_message(
     db: DbSession,
     user: Annotated[User, Depends(require_state_change)],
     request: Request,
+    background_tasks: BackgroundTasks,
     body: MessageIn,
     idempotency_key: Annotated[str, Header(alias="Idempotency-Key")],
 ) -> MessageCreatedOut | JSONResponse:
@@ -90,6 +91,17 @@ async def post_message(
     if replayed:
         # 重放成功按契约返回 200 与原资源（避开装饰器固定 201）
         return JSONResponse(status_code=200, content=json.loads(payload.model_dump_json()))
+    if run is not None:
+        # 先显式提交：确保 Run 在后台任务执行前已落库（与依赖清理顺序解耦）
+        await db.commit()
+        from app.gateway import execute_run
+
+        background_tasks.add_task(
+            execute_run,
+            request.app.state.session_factory,
+            run.id,
+            transport=getattr(request.app.state, "gw_transport", None),
+        )
     return payload
 
 
@@ -111,23 +123,40 @@ async def run_events(
     if run.personal_domain_id != domain.id:
         raise HTTPException(status.HTTP_404_NOT_FOUND, "Run 不存在")
 
+    import asyncio as _asyncio
+
     from app.models import RunEvent
 
-    events = (
-        await db.execute(
-            select(RunEvent)
-            .where(RunEvent.run_id == run.id)
-            .order_by(RunEvent.sequence.asc())
-        )
-    ).scalars().all()
+    terminal = {"completed", "failed", "cancelled"}
 
     async def event_stream() -> AsyncIterator[str]:
-        for event in events:
-            payload = json.dumps(
-                {"type": event.type, "data": event.data, "sequence": event.sequence},
-                ensure_ascii=False,
+        last_seq = -1
+        deadline = _asyncio.get_event_loop().time() + 30.0
+        while True:
+            rows = list(
+                (
+                    await db.execute(
+                        select(RunEvent)
+                        .where(RunEvent.run_id == run.id, RunEvent.sequence > last_seq)
+                        .order_by(RunEvent.sequence.asc())
+                    )
+                ).scalars().all()
             )
-            yield f"data: {payload}\n\n"
+            for event in rows:
+                payload = json.dumps(
+                    {"type": event.type, "data": event.data, "sequence": event.sequence},
+                    ensure_ascii=False,
+                )
+                yield f"data: {payload}\n\n"
+                last_seq = event.sequence
+            if not rows:
+                await db.rollback()  # 释放快照，下一轮可见新写入
+                await db.refresh(run)
+                if run.status in terminal:
+                    break
+                if _asyncio.get_event_loop().time() > deadline:
+                    break
+                await _asyncio.sleep(0.05)
 
     return StreamingResponse(
         event_stream(),