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/video/inference_aggregator.py

@@ -0,0 +1,93 @@
+"""时间窗聚合：按 ``consumable_vision_window_sec`` 桶内众数投票，产出 ``WindowInferenceReady``。
+
+从 ``CameraSessionManager._camera_worker`` 的时间窗计票逻辑独立出来，便于单测。
+消耗 TSV / 终端 Markdown 在 ``VisionClassificationHandler`` 中按「自动确认 / 待确认」分支写入，
+避免待确认事件在日志中先记成具体耗材名。
+"""
+
+from __future__ import annotations
+
+import time
+from dataclasses import dataclass
+
+from app.config import Settings
+from app.services.consumable_vision_algorithm import (
+    ClsTop3,
+    PredictionResult,
+    cls_top3_to_prediction_result,
+    window_bucket_to_best_snap,
+)
+from app.services.video.session_registry import (
+    CameraStreamInferState,
+    SurgerySessionState,
+)
+
+
+@dataclass(frozen=True)
+class WindowInferenceReady:
+    """单个已完成时间窗：原始 top3 快照 + 分类结果 + 墙钟区间（与 monotonic 窗对齐）。"""
+
+    best: ClsTop3
+    prediction: PredictionResult
+    wall_lo: float
+    wall_hi: float
+
+
+class WindowInferenceAggregator:
+    """负责把单路相机的推理快照按时间窗分桶，并产出「桶内最佳」结果。
+
+    本类无状态：状态保存在 ``SurgerySessionState.camera_infer`` 中，
+    便于与原逻辑保持一致；调用方在持有 ``state.lock`` 时调用下面的方法。
+    """
+
+    def __init__(self, *, settings: Settings) -> None:
+        self._s = settings
+
+    def ingest_snapshot_and_collect_ready(
+        self,
+        *,
+        surgery_id: str,
+        camera_id: str,
+        snap: ClsTop3,
+        state: SurgerySessionState,
+    ) -> list[WindowInferenceReady]:
+        """摄入一条推理快照，返回本次因桶满而产出的窗口列表。
+
+        调用方必须已持有 ``state.lock``。
+        """
+        _ = surgery_id
+        _ = camera_id
+        wsec = self._s.consumable_vision_window_sec
+        ready: list[WindowInferenceReady] = []
+        cis = state.camera_infer.setdefault(camera_id, CameraStreamInferState())
+        if cis.stream_t0 is None:
+            cis.stream_t0 = time.monotonic()
+            cis.stream_wall_start = time.time()
+        t_rel = time.monotonic() - cis.stream_t0
+        cis.votes.append((t_rel, snap.t1_name, snap))
+        current_b = int(t_rel // wsec)
+        while cis.next_bucket < current_b:
+            b = cis.next_bucket
+            cis.next_bucket += 1
+            lo, hi = b * wsec, (b + 1) * wsec
+            bucket_pts = [(p, sn) for (t, p, sn) in cis.votes if lo <= t < hi]
+            cis.votes = [
+                (t, p, sn) for (t, p, sn) in cis.votes if not (lo <= t < hi)
+            ]
+            if not bucket_pts:
+                continue
+            best = window_bucket_to_best_snap(bucket_pts)
+            if best is None or cis.stream_wall_start is None:
+                continue
+            wall_lo = cis.stream_wall_start + lo
+            wall_hi = cis.stream_wall_start + hi
+            pred = cls_top3_to_prediction_result(best)
+            ready.append(
+                WindowInferenceReady(
+                    best=best,
+                    prediction=pred,
+                    wall_lo=wall_lo,
+                    wall_hi=wall_hi,
+                )
+            )
+        return ready