feat: 消费停录汇总与查结果同口径，并更新分类与 TSV

- 停录时由 details 经 build_consumption_summary 写 TSV/终端汇总，移除 consumption_log_totals 与时间窗累加
- 更新 consumable_classifier 权重、YAML 标签与 consumable_vision_algorithm
- 扩展 consumption TSV 相关测试与配置注释

Made-with: Cursor

app/services/consumable_vision_algorithm.py

@@ -5,8 +5,9 @@
 
 from __future__ import annotations
 
-from typing import Any
+import functools
 import os
+from typing import Any
 import sys
 from collections import Counter
 from dataclasses import dataclass
@@ -126,6 +127,38 @@ def load_name_to_label_id_from_yaml(path: Path) -> dict[str, str]:
     return out
 
 
+def load_index_to_label_id_from_yaml(path: Path) -> dict[int, str]:
+    """与 ``label_id`` 段：类索引 -> 业务 id 字符串；类名与 YAML 略有不一致时仍可落盘到正确 id。"""
+    try:
+        raw = path.read_text(encoding="utf-8")
+    except OSError:
+        return {}
+    try:
+        data: Any = yaml.safe_load(raw)
+    except yaml.YAMLError:
+        return {}
+    if not isinstance(data, dict):
+        return {}
+    label_raw = data.get("label_id")
+    if not isinstance(label_raw, dict):
+        return {}
+    out: dict[int, str] = {}
+    for k, v in label_raw.items():
+        try:
+            i = int(k)
+        except (TypeError, ValueError):
+            continue
+        if v is None or (isinstance(v, str) and not str(v).strip()):
+            continue
+        out[i] = str(v).strip()
+    return out
+
+
+@functools.lru_cache(maxsize=8)
+def _cached_index_to_label_id(path_resolved: str, mtime_ns: int) -> dict[int, str]:
+    return load_index_to_label_id_from_yaml(Path(path_resolved))
+
+
 def list_sorted_class_names_from_yaml(path: Path) -> list[str]:
     """自 ``names`` 段按类索引升序取类名字符串（与训练/权重一致）。"""
     try:
@@ -218,7 +251,11 @@ def _probs_data_to_numpy1d(raw) -> np.ndarray:
 
 
 def cls_top3_from_result(
-    cls: YOLO, r, name_to_code: dict[str, str]
+    cls: YOLO,
+    r,
+    name_to_code: dict[str, str],
+    *,
+    index_to_label_id: dict[int, str] | None = None,
 ) -> ClsTop3 | None:
     pr = r[0].probs
     if pr is None:
@@ -244,6 +281,7 @@ def cls_top3_from_result(
 
     n2 = n3 = ""
     c2 = c3 = 0.0
+    i2 = i3 = -1
     if len(t5i) > 1:
         i2 = int(t5i[1])
         n2 = str(cls.names.get(i2, "")).strip()
@@ -253,12 +291,19 @@ def cls_top3_from_result(
         n3 = str(cls.names.get(i3, "")).strip()
         c3 = _conf_for_idx(i3)
 
-    def _pid(label: str) -> str:
+    idx_extras = index_to_label_id or {}
+
+    def _pid(label: str, class_idx: int) -> str:
         lb = (label or "").strip()
         if not lb:
             return ""
         norm = _norm_product_name(lb)
-        return (name_to_code.get(norm) or name_to_code.get(lb) or "").strip()
+        c = (name_to_code.get(norm) or name_to_code.get(lb) or "").strip()
+        if c:
+            return c
+        if class_idx >= 0 and class_idx in idx_extras:
+            return idx_extras[class_idx]
+        return ""
 
     return ClsTop3(
         t1_name=n1,
@@ -267,9 +312,9 @@ def cls_top3_from_result(
         t2_conf=c2,
         t3_name=n3,
         t3_conf=c3,
-        t1_pid=_pid(n1),
-        t2_pid=_pid(n2),
-        t3_pid=_pid(n3),
+        t1_pid=_pid(n1, t1i),
+        t2_pid=_pid(n2, i2),
+        t3_pid=_pid(n3, i3),
     )
 
 
@@ -475,12 +520,31 @@ class ConsumableVisionAlgorithmService:
         except Exception as exc:
             raise PredictionError(f"耗材分类推理失败: {exc}") from exc
 
-        snap = cls_top3_from_result(cls_model, r, name_to_code)
+        yp = Path(self._s.consumable_classifier_labels_yaml_path).expanduser()
+        if yp.is_file():
+            st = yp.stat()
+            index_to_label_id = _cached_index_to_label_id(
+                str(yp.resolve()), st.st_mtime_ns
+            )
+        else:
+            index_to_label_id = {}
+
+        snap = cls_top3_from_result(
+            cls_model,
+            r,
+            name_to_code,
+            index_to_label_id=index_to_label_id,
+        )
         if snap is None:
             return None
         if snap.t1_conf < self._s.consumable_min_cls_confidence:
             return None
         pname = snap.t1_name
-        if not pname or pname not in whitelist:
+        if not pname:
             return None
-        return snap
+        pnorm = _norm_product_name(pname)
+        if pnorm in whitelist or pname in whitelist:
+            return snap
+        if (snap.t1_pid or "").strip():
+            return snap
+        return None

app/services/consumption_tsv_log.py

@@ -1,6 +1,6 @@
-"""每例手术一个文本文件（制表符列）：`start_surgery` 时截断并写表头，每次时间窗识别**追加**一行（仅 item_id, item_name, qty, doctor_id, timestamp）。终端 Markdown 时间戳为可读形式；落盘时间戳为 ISO 区间便于程序解析。
+"""每例手术一个文本文件（制表符列）：`start_surgery` 时截断并写表头，每次时间窗识别**追加**一行（仅 item_id, item_name, qty, doctor_id, timestamp）。待确认行首列为 ``pending:{confirmation_id}``，语音落锤后**整行替换**为与客户端一致的最终真值，不再重复追加。终端 Markdown 时间戳为可读形式；落盘时间戳为 ISO 区间便于程序解析。
 
-手术结束时再追加一节汇总行：item_id, item_name, qty（无其它列）。
+手术结束时再追加一节汇总行：item_id, item_name, qty（无其它列）；与 HTTP ``summary`` 同算法，由内存 ``details`` 经 ``build_consumption_summary`` 得到，非录制过程中按窗累计。
 
 时间戳：在拉流起点记录 `time.time()`，与 `time.monotonic()` 时间窗对齐。直播 RTSP 经 OpenCV 一般无可靠绝对时码，以本机接收时刻为准。
 """
@@ -353,25 +353,15 @@ def append_consumption_pending_window(
         )
 
 
-def append_consumption_voice_resolution_line(
+def _build_voice_resolved_tsv_data_line(
     *,
-    surgery_id: str,
     name_to_code: dict[str, str],
     chosen_label: str,
     doctor_id: str,
     wall_epoch: float,
-    tsv_enabled: bool | None = None,
-) -> None:
-    """语音确认后追加一行最终耗材（医生 ID 多为 voice）；top2/3 空列。
-
-    待确认流程下，时间窗仅记「待确认」，此处写入医生选定后的正式记录。
-    """
-    en = settings.consumption_tsv_log_enabled if tsv_enabled is None else tsv_enabled
-    if not en:
-        return
+) -> str:
+    """与客户端一致的最终行（top2/3 空列，timestamp 为单点）。"""
     lb = (chosen_label or "").strip()
-    if not lb:
-        return
     norm = _norm_product_name(lb)
     p = (
         name_to_code.get(norm) or name_to_code.get(lb) or ""
@@ -387,7 +377,7 @@ def append_consumption_voice_resolution_line(
         t2_pid="",
         t3_pid="",
     )
-    line = build_tsv_line(
+    return build_tsv_line(
         name_to_code=name_to_code,
         best=snap,
         doctor_id=doctor_id,
@@ -395,6 +385,98 @@ def append_consumption_voice_resolution_line(
         wall_start_epoch=wall_epoch,
         wall_end_epoch=wall_epoch,
     )
+
+
+def _data_line_starts_with_pending_id(line: str, confirmation_id: str) -> bool:
+    s = (line or "").rstrip("\r\n")
+    if not s or s.startswith("item_id\t"):
+        return False
+    want = f"pending:{(confirmation_id or '').strip()}"
+    first = s.split("\t", 1)[0]
+    return first == _encode_cell(want) or first == want
+
+
+def replace_pending_line_with_voice_resolution(
+    *,
+    surgery_id: str,
+    confirmation_id: str,
+    name_to_code: dict[str, str],
+    chosen_label: str,
+    doctor_id: str,
+    wall_epoch: float,
+    tsv_enabled: bool | None = None,
+) -> None:
+    """将同一 ``confirmation_id`` 下先前追加的「待确认」行整行改为最终真值，不再多追加一行。
+
+    未找到 ``pending:{confirmation_id}`` 时回退为追加（兼容旧文件或行缺失）。读改写全程持
+    :data:`_lock`，避免与并发的 append 交错。
+    """
+    en = settings.consumption_tsv_log_enabled if tsv_enabled is None else tsv_enabled
+    if not en:
+        return
+    if not (chosen_label or "").strip() or not (confirmation_id or "").strip():
+        return
+    new_line = _build_voice_resolved_tsv_data_line(
+        name_to_code=name_to_code,
+        chosen_label=chosen_label,
+        doctor_id=doctor_id,
+        wall_epoch=wall_epoch,
+    )
+    path = resolved_consumption_log_path(surgery_id)
+    with _lock:
+        if not path.is_file():
+            path.parent.mkdir(parents=True, exist_ok=True)
+            with path.open("w", encoding="utf-8") as f:
+                f.write(HEADER)
+                f.write(new_line)
+            logger.warning(
+                "consumption TSV 无文件，无法替换 pending 行，已写表头+最终行: {}",
+                path,
+            )
+            return
+        text = path.read_text(encoding="utf-8")
+        lines = text.splitlines(keepends=True)
+        replaced = False
+        out: list[str] = []
+        for line in lines:
+            if (
+                not replaced
+                and _data_line_starts_with_pending_id(line, confirmation_id)
+            ):
+                out.append(new_line)
+                replaced = True
+            else:
+                out.append(line)
+        if not replaced:
+            logger.warning(
+                "未在 TSV 中找到 pending 行，回退为追加: surgery={} confirm={}",
+                surgery_id,
+                confirmation_id,
+            )
+            with path.open("a", encoding="utf-8") as f:
+                f.write(new_line)
+            return
+        path.write_text("".join(out), encoding="utf-8")
+
+
+def append_consumption_voice_resolution_line(
+    *,
+    surgery_id: str,
+    name_to_code: dict[str, str],
+    chosen_label: str,
+    doctor_id: str,
+    wall_epoch: float,
+    tsv_enabled: bool | None = None,
+) -> None:
+    """已弃用：请使用 ``replace_pending_line_with_voice_resolution`` 并传 ``confirmation_id``。"""
+    if not (chosen_label or "").strip():
+        return
+    line = _build_voice_resolved_tsv_data_line(
+        name_to_code=name_to_code,
+        chosen_label=chosen_label,
+        doctor_id=doctor_id,
+        wall_epoch=wall_epoch,
+    )
     append_consumption_tsv_line(surgery_id, line)
 
 
@@ -467,19 +549,9 @@ class ConsumptionTsvWriter:
         camera_id: str,
         wall_start_epoch: float,
         wall_end_epoch: float,
-        running_totals: dict[str, tuple[str, int]] | None = None,
     ) -> None:
         if not self._s.consumption_tsv_log_enabled and not self._s.consumption_log_markdown_terminal:
             return
-        _tsv_id, totals_key = resolve_consumption_ids(
-            best.t1_name, best.t1_pid, name_to_code
-        )
-        iname = (best.t1_name or "").strip()
-        if running_totals is not None:
-            if totals_key not in running_totals:
-                running_totals[totals_key] = (iname, 0)
-            prev_name, q = running_totals[totals_key]
-            running_totals[totals_key] = (prev_name, q + 1)
         if self._s.consumption_tsv_log_enabled:
             line = build_tsv_line(
                 name_to_code=name_to_code,
@@ -549,19 +621,9 @@ def append_consumption_window(
     camera_id: str,
     wall_start_epoch: float,
     wall_end_epoch: float,
-    running_totals: dict[str, tuple[str, int]] | None = None,
 ) -> None:
     if not settings.consumption_tsv_log_enabled and not settings.consumption_log_markdown_terminal:
         return
-    _tsv_id, totals_key = resolve_consumption_ids(
-        best.t1_name, best.t1_pid, name_to_code
-    )
-    iname = (best.t1_name or "").strip()
-    if running_totals is not None:
-        if totals_key not in running_totals:
-            running_totals[totals_key] = (iname, 0)
-        prev_name, q = running_totals[totals_key]
-        running_totals[totals_key] = (prev_name, q + 1)
     if settings.consumption_tsv_log_enabled:
         line = build_tsv_line(
             name_to_code=name_to_code,

app/services/video/classification_handler.py

@@ -7,7 +7,7 @@
 - 置信度 ≥ 自动阈值但 Top1 不在候选内 → 视 voice_confirmation_enabled 入 pending。
 - 中等置信度 → 入 pending（若有可展示候选项）。
 
-需医生确认时：消耗 TSV / 内存明细记「待确认」（不写模型 top1 商品名）；语音确认后再落最终耗材并更新汇总。
+需医生确认时：消耗 TSV / 内存明细记「待确认」（不写模型 top1 商品名）；语音确认后**替换**该条 TSV 为最终耗材。停录时 TSV/终端汇总与查结果 API 同口径，由 details 经 ``build_consumption_summary`` 生成。
 """
 
 from __future__ import annotations
@@ -85,7 +85,6 @@ class VisionClassificationHandler:
             camera_id=camera_id,
             wall_start_epoch=ready.wall_lo,
             wall_end_epoch=ready.wall_hi,
-            running_totals=state.consumption_log_totals,
         )
 
     async def handle(
@@ -99,7 +98,8 @@ class VisionClassificationHandler:
     ) -> None:
         conf = cls_res.confidence
         label = (cls_res.label or "").strip()
-        item_id = resolve_consumption_item_id(label, "", state.name_to_code)
+        t1_pid = (ready.best.t1_pid if ready is not None else "")
+        item_id = resolve_consumption_item_id(label, t1_pid, state.name_to_code)
         voice_floor = self._s.video_voice_confirm_min_confidence
         if conf < voice_floor:
             return

app/services/video/session_manager.py

@@ -29,6 +29,7 @@ from app.services.video.session_registry import (
 )
 from app.services.video.stream_worker import CameraStreamWorker, redact_rtsp_url
 from app.services.video.types import VideoBackendKind
+from app.schemas import SurgeryConsumptionDetail, build_consumption_summary
 from app.services.consumption_tsv_log import (
     append_consumption_log_summary,
     init_consumption_log_file,
@@ -214,11 +215,21 @@ class CameraSessionManager:
             if isinstance(res, BaseException):
                 logger.warning("surgery task finished with error: {}", res)
 
-        totals = dict(run.state.consumption_log_totals)
+        details = list(run.state.details)
+        detail_rows = [
+            SurgeryConsumptionDetail(
+                item_id=r.item_id,
+                item_name=r.item_name,
+                qty=r.qty,
+                doctor_id=r.doctor_id,
+                timestamp=r.timestamp,
+            )
+            for r in details
+        ]
+        summ = build_consumption_summary(detail_rows)
+        totals = {s.item_id: (s.item_name, s.total_quantity) for s in summ}
         append_consumption_log_summary(surgery_id, totals)
         print_consumption_summary_markdown(totals)
-
-        details = list(run.state.details)
         await self._archive.persist_or_archive(surgery_id, details)
 
     # ------------------------------------------------------------------

app/services/video/session_registry.py

@@ -22,8 +22,7 @@ from app.services.consumable_vision_algorithm import (
     _norm_product_name,
 )
 from app.services.consumption_tsv_log import (
-    append_consumption_voice_resolution_line,
-    resolve_consumption_ids,
+    replace_pending_line_with_voice_resolution,
     resolve_consumption_item_id,
 )
 from app.services.voice_confirm import build_prompt_text
@@ -74,8 +73,6 @@ class SurgerySessionState:
     last_asr_text: str | None = None
     #: 最近一次语音确认错误说明（ASR/解析失败等）。
     last_voice_error: str | None = None
-    #: 视觉时间窗落盘用量累计，供停录时写汇总（item_id -> 首次名称, 次数）。
-    consumption_log_totals: dict[str, tuple[str, int]] = field(default_factory=dict)
 
 
 @dataclass
@@ -281,6 +278,7 @@ class SurgerySessionRegistry:
                 self._finalize_voice_confirmed_consumption_log(
                     state=st,
                     surgery_id=surgery_id,
+                    confirmation_id=confirmation_id,
                     chosen_label=label,
                 )
             try:
@@ -295,20 +293,16 @@ class SurgerySessionRegistry:
         *,
         state: SurgerySessionState,
         surgery_id: str,
+        confirmation_id: str,
         chosen_label: str,
     ) -> None:
-        """待确认流程在语音落锤后：汇总 +1 最终耗材，并追加 TSV 正式行。"""
+        """待确认流程在语音落锤后：将 TSV 中原 pending 行替换为最终真值。停录汇总与 HTTP 一致，由 details 经 ``build_consumption_summary`` 得到。"""
         cl = (chosen_label or "").strip()
         if not cl:
             return
-        _, key_chosen = resolve_consumption_ids(cl, "", state.name_to_code)
-        tot = state.consumption_log_totals
-        if key_chosen not in tot:
-            tot[key_chosen] = (cl, 0)
-        nm, q = tot[key_chosen]
-        tot[key_chosen] = (nm, q + 1)
-        append_consumption_voice_resolution_line(
+        replace_pending_line_with_voice_resolution(
             surgery_id=surgery_id,
+            confirmation_id=confirmation_id,
             name_to_code=state.name_to_code,
             chosen_label=cl,
             doctor_id=self._s.video_voice_confirm_doctor_id,