feat: 手术视频消耗、待确认与持久化改造

- 新增 Alembic 初始迁移、领域明细模型及归档持久化与重试链路\n- 拆分视频会话注册表、分类处理、推理时间窗聚合与流处理\n- 消耗日志：TSV/Markdown 含 top2/top3；item_id 优先产品编码；待确认记「待确认」行，语音确认后落正式行并更新汇总\n- 待确认时内存/DB 明细为占位行，确认后替换；拒绝时移除占位\n- 分类 probs 先 detach/cpu 再转 NumPy，修复 MPS/CUDA 上推理被静默跳过\n- 补充集成测试、归档与设备张量等单测

Made-with: Cursor

app/services/pending_confirmation_port.py

@@ -0,0 +1,54 @@
+"""语音确认服务访问会话状态的端口协议。
+
+把 `VoiceConfirmationService` 对 `CameraSessionManager` 的强依赖解耦为
+`PendingConfirmationStore` 协议；便于单元测试用 fake，并为后续拆分会话管理器
+（`SurgerySessionRegistry` 等）保留切换点。
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Protocol, runtime_checkable
+
+if TYPE_CHECKING:
+    from app.services.video.session_manager import PendingConsumableConfirmation
+
+
+@runtime_checkable
+class PendingConfirmationStore(Protocol):
+    """语音确认链路需要的最小会话接口。"""
+
+    def get_pending_confirmation_by_id(
+        self,
+        surgery_id: str,
+        confirmation_id: str,
+    ) -> "PendingConsumableConfirmation | None":
+        ...
+
+    def get_surgery_candidate_consumables(self, surgery_id: str) -> list[str]:
+        ...
+
+    async def record_voice_parse_failure(
+        self,
+        surgery_id: str,
+        confirmation_id: str,
+    ) -> tuple[int, int]:
+        ...
+
+    async def resolve_pending_confirmation(
+        self,
+        surgery_id: str,
+        confirmation_id: str,
+        *,
+        chosen_label: str | None,
+        rejected: bool,
+    ) -> None:
+        ...
+
+    def record_voice_trace(
+        self,
+        surgery_id: str,
+        *,
+        asr_text: str | None,
+        error: str | None,
+    ) -> None:
+        ...