界面：三维可选「深度响应 / 光流幅值」上色（默认深度），并加一个「诊断信息」开关收拾面板脚下的文字

- What（颜色来源）: `DisplaySettings` 增 `field`（`"depth"` 默认 / `"flow"`），随显示设置持久化
  （`scene3d/prefs.py`：新增 `FIELD_CHOICES`/`DEFAULT_FIELD`，手改文件里的非法值修回默认）。
  `FingerFrameAdapter` 按所选字段先试、失败再用另一个，并把"换了量"这件事记在
  `last_substitution`（新诊断码 `field_substituted`，info 级）；两个字段都拿不到时，
  `last_diagnostic` 取**信息量最大**的那条（域不符/过期/未声明 > 没这个字段），而不是"所选字段"的
  那条空话。设置对话框新增「颜色来源」下拉，下面一行 `field_note` 直接写"这一帧画的是 X，不是所选的 Y"；
  图例（`legend_label`）永远标注**实际**画出来的量，所以设定与画面不一致时两处都能看出来。
  深度响应 `[1]`、光流幅值 `[px]`，两者都是 `approximate`——代码与文案都不把它们说成毫米。
- What（诊断信息开关）: 视图菜单新增 `诊断信息`（可勾选、持久化 `show_diagnostics`、**默认关**，
  由 `DisplayPrefsController` 持有并推给两个面板）。关掉后：2D 面板只留"画面 + 计划说明 + 深度色标"，
  三条解释流水线的标签（箭头按位移放大、原始图像 → 算法输出、全部帧）隐藏；3D 面板只留
  色条 + 一行图例 + 来源，info 级诊断隐藏。**永不隐藏**：warning/error 级诊断（测量是否过期、
  字段是否本帧的、是否与设备核对过——`geometry_unverified` 虽然是 info 也白名单保留）与
  深度色标。3D 图例改成 `ShrinkableLabel`（一行、裁切 + tooltip，不再折行），色标数值从 4 位有效数字
  收到 3 位（"0.0334 – 0.899"）。设置对话框控件宽度 88 → 108 px（不再需要迁就 240 px 面板）。
- What（取景脚本）: `tools/ui_snapshot.py --scene flow3d` 现在同时给 flow 与 depth（各带 provenance），
  所以「颜色来源」开关有东西可切；再加 `.repro/shot_variants.py` 一次出四张图
  （深度版 / 光流版 / 诊断开 / 设置对话框）。
- What（测试）: `test_scene3d_adapters.py` 增 3 项（默认深度、换量时给出 `field_substituted` 并在恢复后
  清掉、depth 不在场景栅格时按 `field_domain_mismatch` 拒绝而不是缩放）；`test_scene3d_view.py` 增
  1 项（field 的持久化与修复，并把 `field in FIELD_CHOICES` 加进"手改文件会被修复"的参数化）；
  `test_scene3d_pane_ui.py` 增 3 项（两种量可选 + 说明行 + 恢复的偏好会移动下拉；诊断开关只隐藏 info、
  保留 warning 与"未核对"；2D 脚注三标签的显隐与窗口开关一致）；`test_display_prefs.py` 增 1 项
  （开关推到两个面板、写偏好）。`test_ui_smoke.py` 与 `tools/ui_check.py` 里"箭头图例可见"的四条断言
  改为按新规则检查（默认隐藏、开诊断后可见、关掉再隐藏）。
- Why: ①"深度"比"光流"更贴近操作员口中的"压得多深"，但两者都只是近似量，所以做成开关而不是替换——
  并且换量必须说出来（否则画面会被读成所选的量）。②面板脚下的解释文字是"内部管线"而不是"画面是什么"，
  拍视频与日常看都只需要后者；用一个开关收起，比做一个滚动日志面板诚实（这些是每帧状态，不是事件）。
- Verification: pytest shared/tests apps/studio/tests（offscreen）658 passed / 0 failed / 15 skipped；
  真 GL 组（gl_smoke + view + pane_ui）70 passed / 0 failed / 1 skipped；
  `tools/ui_check.py` 590 项检查 0 失败；`.repro/locale_check.py` 0 缺失（新增 7 个 key 双语齐备）。
  目视验收：同一批合成数据渲染的四张图（深色主题）——深度版（默认）、光流版、诊断打开、设置对话框。
- Not in this change: 不改采集/记录/契约之外的三维渲染；不做页签容器与窄窗口回退；曲线量程仍是出厂 ±20 N。
  真机待验证：手指路径上 `packet.depth` 的栅格是否就是场景的 300×280（若不是，面板会写明并回退到光流）。

apps/studio/tests/test_scene3d_adapters.py

@@ -73,6 +73,84 @@ def test_the_flow_magnitude_handles_both_shapes():
     assert flow_magnitude(FakePacket(flow=np.zeros((2, 2, 3), dtype=np.float32))) is None
 
 
+def test_the_finger_adapter_prefers_the_quantity_the_operator_chose(scene):
+    """The choice between the two quantities is a display setting, honoured per frame.
+
+    The depth response is what an operator reads as "how deep", so it is the default —
+    and when the frame carries it, that is what is drawn.
+    """
+    adapter = FingerFrameAdapter(
+        geometry_id=scene.geometry_id,
+        field_domain_id=scene.field_domain_id,
+        shape_hw=scene.shape_hw,
+    )
+    assert adapter.preferred_field == "depth"
+    depth = np.full(scene.shape_hw, 2.5, dtype=np.float32)
+    flow = np.zeros((*scene.shape_hw, 2), dtype=np.float32)
+    flow[..., 0] = 7.0  # |flow| = 7 px
+
+    frame = adapter.snapshot(FakePacket(flow=flow, depth=depth), runtime=None, identity=_identity())
+    assert frame is not None
+    assert frame.scalar_descriptor == DEPTH_RESPONSE
+    assert frame.scalar_descriptor.unit == "1"
+    assert np.allclose(np.asarray(frame.scalar), 2.5)
+    assert adapter.last_substitution is None, "the chosen field was drawn; nothing to say"
+
+    adapter.set_preferred_field("flow")
+    frame = adapter.snapshot(FakePacket(flow=flow, depth=depth), runtime=None, identity=_identity())
+    assert frame is not None
+    assert frame.scalar_descriptor == FLOW_MAGNITUDE
+    assert np.allclose(np.asarray(frame.scalar), 7.0)
+    assert adapter.last_substitution is None
+
+
+def test_a_frame_that_cannot_honour_the_choice_says_which_quantity_it_drew(scene):
+    """Falling back is allowed; falling back silently is not.
+
+    The frame is still a measurement and the legend names it, but the *setting* said
+    something else, and the pane has to be able to say so instead of letting the picture
+    read as the quantity the operator asked for.
+    """
+    adapter = FingerFrameAdapter(
+        geometry_id=scene.geometry_id,
+        field_domain_id=scene.field_domain_id,
+        shape_hw=scene.shape_hw,
+    )
+    # No depth in this frame: the flow is all it has.
+    flow = np.zeros((*scene.shape_hw, 2), dtype=np.float32)
+    flow[..., 0] = 4.0  # |flow| = 4 px
+    frame = adapter.snapshot(FakePacket(flow=flow), runtime=None, identity=_identity())
+    assert frame is not None
+    assert frame.scalar_descriptor == FLOW_MAGNITUDE
+    note = adapter.last_substitution
+    assert note is not None and note.code == "field_substituted"
+    assert note.details["wanted"] == "depth" and note.details["shown"] == "flow"
+    assert "no depth response" in note.details["why"]
+
+    # ...and the note goes away as soon as the chosen field can be drawn again.
+    depth = np.full(scene.shape_hw, 1.0, dtype=np.float32)
+    frame = adapter.snapshot(
+        FakePacket(flow=flow, depth=depth), runtime=None, identity=_identity()
+    )
+    assert frame is not None and frame.scalar_descriptor == DEPTH_RESPONSE
+    assert adapter.last_substitution is None
+
+
+def test_a_depth_response_on_another_grid_is_refused_not_resized(scene):
+    """The depth field gets the same treatment as the flow: no resizing, ever."""
+    adapter = FingerFrameAdapter(
+        geometry_id=scene.geometry_id,
+        field_domain_id=scene.field_domain_id,
+        shape_hw=scene.shape_hw,
+    )
+    wrong = np.zeros((7, 7), dtype=np.float32)
+    frame = adapter.snapshot(FakePacket(depth=wrong), runtime=None, identity=_identity())
+    assert frame is None, "a depth response that is not on the scene's grid was coloured"
+    item = adapter.last_diagnostic
+    assert item is not None and item.code == "field_domain_mismatch"
+    assert item.details["depth"] == "7x7"
+
+
 # ----------------------------------------------------------------- finger adapter
 def test_the_finger_adapter_publishes_a_field_already_in_the_domain(scene):
     adapter = FingerFrameAdapter(