From 1aa454311fa44d3ab8c52846275f50231a1879ed Mon Sep 17 00:00:00 2001
From: ykcheungab <ykcheungab@connect.ust.hk>
Date: Sun, 4 Oct 2026 16:04:02 +0800
Subject: [PATCH] Make the cached analysis views reachable over a recording
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Replay moved the cursor onto the result's timeline, so the recording's own
picture was not offered at all and ``use_recording()`` had no caller. Adopting a
result now keeps the cursor on the recording, opens the panel on a derived view,
and serves scrubs, playback and display changes from the cache (Part B §9).
Verified: apps/studio 455 items / 0 failed / 18 skipped, shared 427 / 0 / 1,
autocal 358, ui_check 628 checks / 0 failures, lint gate clean.
---
 apps/studio/docs/replay-analysis-part-b.md    |  96 ++--
 apps/studio/docs/replay-panel.md              |  15 +-
 .../src/orisys_studio/replay_controller.py    |  90 +++-
 apps/studio/tests/test_replay_analysis.py     |  13 +-
 apps/studio/tests/test_replay_wiring.py       | 507 ++++++++++++++++++
 5 files changed, 649 insertions(+), 72 deletions(-)
 create mode 100644 apps/studio/tests/test_replay_wiring.py

diff --git a/apps/studio/docs/replay-analysis-part-b.md b/apps/studio/docs/replay-analysis-part-b.md
index c44ea10..c5be65e 100644
--- a/apps/studio/docs/replay-analysis-part-b.md
+++ b/apps/studio/docs/replay-analysis-part-b.md
@@ -15,14 +15,18 @@
 | `recording` | 录像自己的 `raw_*.mp4`，按需解码 | 这个画面是原始相机图像：它自己没有光流也没有深度 |
 | `analysis` | 存下来的分析结果，按帧读 | 有 pass 记录下来的光流样本、候选网格与深度图 |
 
-模式说的是**画面从哪里读**，而"有没有可画的东西"是另一件事：只要有结果，二维面板的模式
-选择器就提供 深度图/箭头（时间轴在录像上时连"原始图像"一起提供），选了派生模式的那些帧
-一律从结果读、按窗口唯一的 2D 出口画 —— 绝不会把原始图像当成箭头图交出去。
+模式说的是画面从哪里读，"有没有可画的东西"是另一件事：只要有结果，选择器就提供箭头与深度图
+（连"原始图像"一起），派生模式那些帧一律从结果读、按窗口唯一的 2D 出口画。
 
 图表、二维画布、三维面板都读这一个游标，模式由 `ReplayEngine.mode()` 回答；「从视频重算」
 因此不再是第三种来源，而是**显式重跑**这一遍的开关（第 10 节）。时间基准是**信号时间**
 （Part A 第 5 节），进度、位置、单步与跳转都说它。
 
+**应用里的时间轴永远是录像的。** 引擎留在 `recording`：游标、帧 id 与时长都是录像自己的，
+结果挂在游标后面当缓存（`available_result()`），`_adopt_session()` 调用 `use_recording()` 保证
+这一点。缓存有三个触点，不必先打开「从视频重算」：复用已有结果、pass 完成、选择器选中派生模式
+（时间轴不动）。
+
 **帧身份只来自录像自己的计数器。** 证据录像 id 是 20…52 而非 0…32，所以
 `AnalysisSession.frame(frame_id)` 对结果里没有的 id 报 `AnalysisNotFoundError`：把索引当 id
 传进去不会被"顺手改成能读的那一帧"，那会返回一张看起来完全正确的图。按时间取帧走
@@ -43,15 +47,14 @@
 
 **并发与工作目录。** 开链路会往当前目录写 2 MB 预览图（Part A 第 6 节），所以 pass 跑在**本
 进程**的一个线程里，启动前要求没有采集会话、没有预览、也没有第二遍分析，并且**本模块从不改
-工作目录**。锁只覆盖开链路那一小段（`_working_directory` 在 `_OpenChain.open()` 内进出），
-逐帧处理在锁之外，GUI 线程这期间正常刷新。关窗时先取消并等待 pass。
+工作目录**。锁只覆盖开链路那一小段，逐帧处理在锁之外，GUI 线程这期间正常刷新。
 
 ## 11. 三视图如何被同一个位置喂
 
 | 视图 | 读什么 | 怎么读 |
 |---|---|---|
 | 图表 | `forces.csv`（记录值） | `charts.set_cursor(t)` 用同一游标；结果里不存力 |
-| 二维画布 | 结果的 `flow`、`depth` 与 `valid_mask` | `ImageFrame(source="analysis")` 走窗口唯一的 2D 出口 `_show_image_frame`，箭头、暗点、配色、域边界与深度视图与实时/重算**同一套代码** |
+| 二维画布 | 结果的 `flow`、`depth` 与 `valid_mask` | `ImageFrame` 走窗口唯一的 2D 出口 `_show_image_frame`（录像时间轴上的派生视图带 `source="replay"` 与 `result_available=True`，因此"原始图像"仍在选择器里），箭头、暗点、配色、域边界与深度视图与实时/重算**同一套代码** |
 | 三维面板 | 同一帧的 `SdkPacket` | `session.packet(frame)` → `Scene3DView.submit_packet`，走同一个 adapter、域掩码 latch 与有效性求交 |
 
 **深度视图也由结果喂。** `AnalysedFrame.depth` 就是第 1 节那幅量化平面重建出来的
@@ -68,20 +71,20 @@
 
 `dx`/`dy` 是**画出来的**位移，已经带上长度余量那一步的缩放：只用前两个因子会让每根箭头短
 20%（finger 余量 0.8），330 个样本里 319 个不同、活动集也不同 —— 实测的，不是推导的。余量补
-上后，重建流场重新过一遍共享选区器，样本集与存下来的**逐元素相等**（33 帧全覆盖）。三维面板
-直接拿存下来的 `FlowSamples`，所以"两张图是同一个测量"由构造决定。
+上后，重建流场重新过一遍共享选区器，样本集与存下来的**逐元素相等**（33 帧全覆盖）。
 
 ## 12. 显示设置是读
 
 阈值、配色范围、暗点开关都不在失效键里（Part A 第 3 节），所以改动它们必须**只重画**：阈值
 经 `AnalysisReader.samples(policy=…)` 作用在存下来的 `magnitude` 上，配色范围与暗点开关只
 存在于绘制调用里。走线是 `Scene3DPanel.settingsChanged` → `on_display_settings()` →
-`session.serve_policy()` → `ReplayEngine.redraw()`，重画不移动游标、不写文件。
+`session.serve_policy()` → `ReplayEngine.redraw()`，重画不移动游标、不写文件。这条走线问的是
+`available_result()` 而不是 `analysis_active()`（结果属于录像、游标在录像的时间轴上）；被拒的
+栅格要由 `policy_divergence()` 报出是哪几项（第 14 节）。
 
 **栅格类设置被拒绝**：`flow_grid_spacing`、`flow_edge_margin`、`max_arrows`、
-`flow_length_margin`、`flow_arrow_scale` 决定候选位置**在哪里**，而记录下来的位置是 pass 当时
-测的那些。`serve_policy()` 返回 `False` 并提示哪一项动了，画面继续按结果自己的栅格绘制 —— 画
-一份"看起来一样、其实采在别处"的箭头才是真正错误的选项。
+`flow_length_margin`、`flow_arrow_scale` 决定候选位置在哪里，而记录下来的位置是 pass 当时测的
+那些；画面继续按结果自己的栅格绘制，而不是画一份"看起来一样、其实采在别处"的箭头。
 
 `app_window.py` 的接线只有三处，供 Part C 使用：`_analysis_service(recording, cancelled, *,
 on_progress)`（`chain_spec_for` 从录像自己的 `session.json` 取链路，因此结果可复现）、
@@ -92,36 +95,25 @@ on_progress)`（`chain_spec_for` 从录像自己的 `session.json` 取链路，
 ## 13. 未完成与取消
 
 `AnalysisSession.__init__` 在 `reader.complete` 为假时抛 `ResultPolicyError`；
-`ResultLibrary.open(..., allow_incomplete=True)` 也拒绝 ——「可查看」与「可回放」是两个不同的
-断言。面板的 `cancelled` 与 `failed` 是两个独立状态，都不会显示成完成；这两种情况下时间轴留
-在 `recording` 模式。
+`ResultLibrary.open(..., allow_incomplete=True)` 同样拒绝：可查看与可回放是两个断言。面板的
+`cancelled` 与 `failed` 是两个独立状态，都不会显示成完成；这两种情况下时间轴留在 `recording`。
 
 ## 14. 验收证据
 
 `apps/studio/tests/test_replay_analysis.py`（15 项，真实素材）：
 
 * **跳转等于顺序播放**：顺序走过全部 33 帧记下每帧，再按 `time_of(frame_id)` 直接跳到 5 个
-  采样位置。容差是零：帧 id、`rows/cols/dx/dy/active` 逐元素相等，流场按位相等（非候选格按
-  构造是 `NaN`，用 `equal_nan` 比较）。跳转精确的原因是给它的时间就是那一帧自己的 `t_rel_s`。
-* **显示设置不写结果**：`scan_results(cache)` 的目录集合在改阈值（0.2 与 1e9）、改显示增益、
-  重复同一栅格前后完全相同。
-* **取消不被当完整**：pass 在第一帧后取消 → 面板报 `cancelled`、`analysis_active()` 为假、
-  引擎停在 `recording` 模式、目录里 `complete=false` 且默认读端抛
-  `AnalysisIncompleteError`。
-* **回归**：`apps/studio/tests` 440 项 / 0 失败 / 18 跳过（原 425 项：新增 15 项，既有断言一条
-  未改）。把分析终态写进状态栏时，
-  `test_the_replay_reset_is_the_display_reset_and_never_the_device` 失败过 —— 合成录像的配置
-  缺失让 pass 在断言之前结束并覆盖了 `复位显示` 的提示；修的是消息落在哪里。
+  采样位置，帧 id、`rows/cols/dx/dy/active` 与流场逐项相等（非候选格按构造是 `NaN`）。
+* **取消不被当完整**：pass 在第一帧后取消 → 面板报 `cancelled`、时间轴留在 `recording`、目录里
+  `complete=false` 且默认读端抛 `AnalysisIncompleteError`。
 
-**深度那一片加的验收**（同一个文件再 +8 项，另加 `shared/tests/test_analysis_depth.py` 10 项）：
+**深度那一片加的验收**（同一文件再 +8 项，另加 `shared/tests/test_analysis_depth.py` 10 项；
+往返误差与尺寸见 Part A 第 1 节）：
 
-* **深度往返**：另跑一遍链路，逐帧比 `depth_map`，误差上界是该帧自己的 `(max − min) / 510`，
-  在 box 录像（33 帧，1 帧非空、76 px）和 264 帧的 `long_8x.mp4`（241 帧空、23 帧非空、
-  最坏 336 px）上都是 0 失败；非零计数、min/max 与链路的完全相等。
-* **跳转等于顺序播放（深度模式）**：把上面那条测试扩到深度 —— 顺序播放记下每帧渲染出的深度
-  图，再跳到 5 个位置，帧 id、`depth`、统计与**渲染出的字节**逐项相等。
-* **尺寸**：空帧 4,048 B（固定 4,042 + 6 B 载荷头），最坏非空帧 4,534 B（+486 B）；
-  按实测比例 4,090 B/帧，十分钟 70.2 MB（上限 81.6 MB）。断言写死了这三个数。
+* **深度往返**：另跑一遍链路逐帧比 `depth_map`，在 box 录像与 264 帧的 `long_8x.mp4` 上都是
+  0 失败；非零计数、min/max 与链路的完全相等。
+* **跳转等于顺序播放（深度模式）**：顺序播放记下每帧渲染出的深度图，再跳到 5 个位置，帧 id、
+  `depth`、统计与渲染出的字节逐项相等。
 * **模式切换不写结果**：四个模式各画一次，`scan_results(cache)` 的目录集合不变。
 * **"没什么可画"只在该说的时候说**：无结果时 `plan_image` 只给 `recorded` 并带
   `replay_raw_note`；有结果时该行消失、四个模式都在，默认仍是原始图像。两种语言各断言一次，
@@ -142,27 +134,41 @@ scratch cache、游标停在存下来的帧 id 上）：
 二维画面是候选网格上的暗点阵（该帧活动样本很少），三维是手指岛面加同一套样本点；"同一个样本
 集"由上面的逐元素断言保证，快照证明这条链路真的走到了 GL。
 
-**深度那一片的快照**（`.scratch/replay-analysis-a2/snapshots/`，同一个窗口、同一个游标、
-同一个模式选择器；时间轴停在录像上，所以"原始图像"是真的相机图，两个派生模式读结果）：
+**深度那一片的快照**（`.scratch/replay-analysis-a2/snapshots/`，同一个窗口、同一个游标、同一个
+选择器；"原始图像"是真的相机图，两个派生模式读结果）：
 
 | 文件 | 内容 |
 |---|---|
 | `analysis-finger-frame-33-2d-raw.png` / `-arrows.png` / `-depth.png` | finger 链路帧 33，深度非零 76 px（峰值 0.0242） |
-| `analysis-box-frame-27-2d-raw.png` / `-arrows.png` / `-depth.png` | box 链路帧 27，`depth_map` 365×365 **恰好全零**，面板显示"本帧深度全为 0（133225 px 中 0 个非零）" |
+| `analysis-box-frame-27-2d-raw.png` / `-arrows.png` / `-depth.png` | box 链路帧 27，`depth_map` 365×365 全零，面板显示"本帧深度全为 0（133225 px 中 0 个非零）" |
 
 同一帧三张图的大小不同（raw 96.6 kB 是相机图，arrows 4.9 kB，depth 4.2 kB），这本身就是
 "三个模式画的是三样东西"的证据。
 
+**接线那一片的验收**（`apps/studio/tests/test_replay_wiring.py`，7 项，真实录像 + 真实 pass；录像
+先复制到临时目录、结果写进它自己的 cache，窗口因此复用而不开链路）：操作员路径断言引擎仍在
+`recording`、`available_result()` 非空、选择器给出"原始图像"与三个派生模式、屏上是复用的结果而
+「从视频重算」是关的（改前报 `'analysis' == 'recording'`；选择器里当时没有"原始图像"）。
+
+箭头与深度的字节分别与存下来的候选/`active` 集合和 `depth_map_to_bgr(存的 depth, …)` 相等
+（这份 box 结果每帧全零，面板用 `panels.depth_empty` 说明）；跳转、单步、真定时器播放（0.4 s）
+与 pass 完成四条都在录像时间轴上读结果。另两项：改阈值只重画、改栅格被拒但画面照画（改前抛
+`TypeError`）。
+
+**接线那一片的快照**（`.scratch/replay-reachable/snapshots/`，真实 GL、真窗口、没有手工调用
+`use_recording()`）：`wiring-operator-frame-40-2d-{raw,arrows,depth}.png` 是同一帧的三个模式
+（86 个活动样本；深度全零），`-playback-2d-arrows.png` 是播放到该时刻抓的，`-window.png` 为整窗。
+
 ## 15. 已知限制
 
 * 结果里没有力：力曲线始终读记录的 `forces.csv`。深度有了（第 1、11 节），但**三维面板仍只
-  拿光流**：`session.packet()` 的 `depth` 依旧是 `None`、`depth_valid` 依旧是 `False`，
-  因为把深度接进三维会改变三维图形的构成，而那不在这一片里。二维深度视图是这一片交付的东西。
-* 结果覆盖的帧是 pass 走过的那些：录像比结果长时游标可以离开结果范围（面板说"这个分析结果里
-  没有可显示的帧"），不会静默重复最后一帧。
+  拿光流**：`session.packet()` 的 `depth` 依旧是 `None`、`depth_valid` 依旧是 `False` —— 接进
+  三维会改变三维图形的构成，不在这一片里。
+* 结果覆盖的帧是 pass 走过的那些：**结果的第一帧可能晚于录像的起点**（box 链路构造时吃掉前
+  7 个解码帧，这份结果从 t=0.4162 s 开始），最后一批帧也可能早于录像的终点。游标落在这两段里
+  时，派生视图按 `frame_id_at()` 的"该时刻或之前、越界夹到端点"给帧：起点之前画的是最早那一帧
+  测量，终点之后画的是最后一帧。两段没有各自的提示，"不会静默重复最后一帧"在本片没有兑现。
 * 栅格类显示设置不生效（第 12 节）；换栅格就是重跑一遍，那是**另一个键**。
 * 一个进程同时只有一遍分析在跑；采集会话运行期间不启动分析。
-* 时间轴与模式选择是两件事，而"有没有结果"是录像的属性、不是游标的属性：
-  `ReplayController.available_result()` 查的是**本录像**在 `ResultLibrary` 里的那一份，所以
-  即使 `ReplayEngine.use_recording()` 让引擎不再借用 session（目前 UI 没有第二条路会走到
-  这个状态），二维面板照样能画箭头和深度，`replay_raw_note` 也不会重新出现。
+* "有没有结果"是录像的属性、不是游标的属性：`available_result()` 查的是本录像在
+  `ResultLibrary` 里的那一份，第 9 节的接线之后这是常态。
diff --git a/apps/studio/docs/replay-panel.md b/apps/studio/docs/replay-panel.md
index ff80496..3af2573 100644
--- a/apps/studio/docs/replay-panel.md
+++ b/apps/studio/docs/replay-panel.md
@@ -5,9 +5,9 @@ capability states its requirement, data source and a falsifying check. The resul
 playback semantics are owned by Part A ([replay-analysis.md](replay-analysis.md)) and Part B
 ([replay-analysis-part-b.md](replay-analysis-part-b.md)).
 
-Status: Parts A and B shipped at `cf49b08`; the depth channel — result format and 2D view — is
-uncommitted; the cache-root and lifetime policy and the display-control separation (Part C)
-are decided below, not implemented.
+Status: Parts A and B shipped at `cf49b08`, the depth channel at `9e20fd0`; the result-aware
+mode plan and the wiring that reaches it are in the working tree (Part B §9, §14). The cache-root
+and lifetime policy and the display-control separation (Part C) are decided below, not implemented.
 
 ## 1. Load a video
 
@@ -110,9 +110,12 @@ frame's depth is all zeros, `panels.depth_unavailable` says the frame has no dep
 all — different facts, neither drawing the processed image.
 
 A mode needing derived data is served from the result, never faked: with none, that view is not
-offered and `panels.replay_raw_note` names the action that produces a result.
+offered and `panels.replay_raw_note` names the action that produces a result. The state above is
+what an operator reaches by opening a recording that has a result: the cursor stays on the
+recording's own timeline and the panel opens on a derived view, the recorded one still offered.
 Verify: `test_image_backend.py` (the plan), `test_replay_analysis.py::test_a_derived_mode_is_served_from_the_result_even_on_the_recording_timeline`,
-`::test_the_empty_depth_note_is_not_the_missing_depth_note`, `test_analysis_depth.py::test_an_empty_map_costs_a_header_and_states_its_shape`.
+`::test_the_empty_depth_note_is_not_the_missing_depth_note`, `test_analysis_depth.py::test_an_empty_map_costs_a_header_and_states_its_shape`,
+`test_replay_wiring.py::test_the_operator_sees_the_cached_views_over_the_recording`.
 
 ## Cache and export policies
 
@@ -151,7 +154,7 @@ possibility, belonging to export rather than the cache. Not implemented.
 |---|---|---|
 | Format, cache, reader, pass | `orisys_app_shared.analysis` | A/B shipped; depth channel uncommitted |
 | Timeline, cursor, policy, pass thread, progress surface | `replay_analysis.py`, `replay.py`, `replay_controller.py`, `analysis_worker.py`, `ui/analysis_progress.py` | shipped; the depth view and the result-aware mode plan are uncommitted |
-| Mode plan, drawing, window wiring | `image_backend.py`, `acquisition/arrows.py`, `app_window.py` | shipped (Part B §12 names the seams); the result-aware plan is uncommitted |
+| Mode plan, drawing, window wiring | `image_backend.py`, `acquisition/arrows.py`, `app_window.py` | shipped (Part B §12 names the seams); the wiring from the operator's open action, and its snapshots, are in the working tree |
 | Cache root and lifetime | `analysis/cache.py`, app preferences | decided, not implemented |
 | Display-control separation: rail reset as the measurement reference, a separate display baseline, the F8 view reset | `toolbar.py`, `display_prefs.py`, `app_window.py` | not started; Part B §12 records these as unchanged. The pass accepts a `reference_frame` but `_analysis_service` passes none |
 
diff --git a/apps/studio/src/orisys_studio/replay_controller.py b/apps/studio/src/orisys_studio/replay_controller.py
index 2379062..4a96734 100644
--- a/apps/studio/src/orisys_studio/replay_controller.py
+++ b/apps/studio/src/orisys_studio/replay_controller.py
@@ -29,6 +29,7 @@ from typing import Any, Protocol
 
 import numpy as np
 from orisys_app_shared.acquisition.image_backend import (
+    ARROW_PURE,
     DERIVED_MODES,
     SOURCE_ANALYSIS,
     SOURCE_REPLAY,
@@ -45,6 +46,7 @@ from orisys_studio.replay_analysis import (
     AnalysisSession,
     ResultLibrary,
     ResultPolicyError,
+    policy_divergence,
 )
 
 logger = logging.getLogger(__name__)
@@ -620,7 +622,13 @@ class ReplayController(QObject):
 
     # ---------------------------------------------------------------- analysis
     def analysis_session(self) -> AnalysisSession | None:
-        """The stored result the timeline is reading, if any."""
+        """Whatever result is open: the cursor's own, else the library's.
+
+        Not the question a view asks — that is :meth:`available_result`, which answers "is
+        there a measurement for *this recording*" and is what the app's timeline (always
+        the recording's) is served from. This one is the raw pair, kept for a caller that
+        wants the session itself.
+        """
         session = self._engine.analysis()
         if session is not None:
             return session
@@ -658,19 +666,25 @@ class ReplayController(QObject):
         ``test_replay_analysis`` that the result directory set is unchanged is a test of
         this method, not of the panel.
 
+        It is asked of ``available_result``, not of ``analysis_active``: a result is a
+        property of the recording, and the cursor is normally on the recording's own
+        timeline while the arrows and the depth map are drawn from it. Gating on the
+        timeline instead would make a threshold change stop mattering in exactly the
+        state this view is served in.
+
         A policy that moves the sample *grid* cannot be served by the stored candidate
-        locations; :meth:`AnalysisSession.serve_policy` says so, the message names the
-        key, and the frame is still redrawn from the result's own grid, because
+        locations; :meth:`AnalysisSession.serve_policy` refuses it, the message names the
+        keys, and the frame is still redrawn from the result's own grid, because
         re-gridding it would draw samples nothing measured.
         """
-        session = self.analysis_session()
-        if session is None or not self.analysis_active():
+        session = self.available_result()
+        if session is None:
             return
         from orisys_app_shared.scene3d.visual_policy import policy_for_settings
 
         policy = policy_for_settings(settings)
-        moved = not session.serve_policy(policy)
-        if moved:
+        if not session.serve_policy(policy):
+            moved = policy_divergence(session.stored_policy(), policy)
             self._host._notify(t("playback.analysis_grid_mismatch", keys="·".join(moved)))
         self._engine.redraw()
 
@@ -765,15 +779,54 @@ class ReplayController(QObject):
         return True
 
     def _adopt_session(self, session: AnalysisSession) -> None:
-        """One timeline: the cursor, the charts and both figures read this result."""
+        """One timeline, one cache behind it: the cursor, the charts and both figures.
+
+        **The engine stays on the recording.** The stored result is adopted *behind* the
+        cursor — :meth:`available_result` answers "is there a measurement for this
+        recording?" — and ``ReplayEngine.use_recording()`` is what keeps the cursor's own
+        timeline the recording's: its signal time, its own frame ids, and its whole span,
+        so a position the pass did not cover is still a position the operator can be at.
+        Moving the cursor onto the result instead (``use_analysis``) is what made the
+        recording's own picture unreachable: with the cursor on a result, the recorded
+        view has nothing to decode.
+        """
         self.results.open_session(session)
-        self._engine.use_analysis(session)
+        self._engine.use_recording()
         self._charts.set_recording_source(self._engine.reader())
-        self._image_panel.set_source_kind("analysis")
+        # The timeline is the recording's, so the panel is told "replay" — the same word
+        # ``apply_source`` uses for an open recording. It used to be told "analysis",
+        # which is not a 2D source kind at all: ``ArrowPanel`` knows live/recompute/replay
+        # and coerced it, so a stored result was worded as the live pipeline.
+        self._image_panel.set_source_kind("replay")
         self._mesh_panel.set_source_kind("analysis")
         if self._actions is not None:
             self._actions.recompute.setEnabled(True)
         self.render_frame()
+        # A result just became readable, so it is what the operator is shown — unless a
+        # derived view was already on screen (a depth view chosen while the pass ran).
+        self.serve_stored_view()
+
+    def serve_stored_view(self) -> bool:
+        """Put the cached view on screen when a result becomes readable.
+
+        The trigger for "draw the cache without being asked": a stored result is adopted
+        — on opening a recording that has one, or when a pass commits — and the panel is
+        showing the recording's own picture, so the operator would otherwise be left
+        looking at the camera while the measurement sits on disk. The recording's own
+        picture stays in the same selector, one step away, and the recompute switch is
+        not involved.
+
+        Answers whether the selection moved. A derived selection is left alone: the mode
+        the operator chose is not this method's to override.
+        """
+        current = str(getattr(self._image_panel, "view_mode", lambda: "")() or "")
+        if current in DERIVED_MODES:
+            return False
+        # The plan for the frame just drawn offers the derived modes (the frame carries
+        # ``result_available``), so this selection is one the panel can honour.
+        self._image_panel.set_view_mode(ARROW_PURE)
+        self.render_frame()
+        return True
 
     def _display_policy(self) -> Any:
         from orisys_app_shared.scene3d.visual_policy import policy_for_settings
@@ -907,6 +960,10 @@ class ReplayController(QObject):
         map can be drawn at all. The session the library holds is checked against the
         recording that is open, so a previous recording's result cannot answer for this
         one.
+
+        This is the normal state of a replay, not a fallback: the cursor stays on the
+        recording's own timeline (``_adopt_session``) and the result is the cache behind
+        it, so the answers and the view modes come from here.
         """
         session = self._engine.analysis()
         if session is not None:
@@ -950,12 +1007,13 @@ class ReplayController(QObject):
     def render_frame(self) -> None:
         """Draw what the 2D panel should show for the current replay cursor.
 
-        One method for both timeline modes, because it is one cursor: in analysis mode
-        the picture is the stored frame's own flow field and depth map (the pass's
-        candidate locations, on the domain grid), and in recording mode it is the
-        decoded video frame. The window's single 2D sink draws both through the *same*
-        backend, so the arrows, the idle dots, the colour range and the domain boundary
-        behave identically in the two modes.
+        One method for both timeline modes, because it is one cursor: with the cursor on
+        a stored result (``analysis_active``) the picture is the stored frame's own flow
+        field and depth map, and with the cursor on the recording's own timeline it is the
+        decoded video frame — unless the *view* the operator asked for is a measurement,
+        in which case it is read from the result for the same cursor. The window's single
+        2D sink draws all of them through the *same* backend, so the arrows, the idle
+        dots, the colour range and the domain boundary behave identically.
 
         **A mode that needs derived data is served from the result, never faked.** The
         arrows and the depth map are the pass's measurements; with a result in hand they
diff --git a/apps/studio/tests/test_replay_analysis.py b/apps/studio/tests/test_replay_analysis.py
index bdaa64b..320ddbd 100644
--- a/apps/studio/tests/test_replay_analysis.py
+++ b/apps/studio/tests/test_replay_analysis.py
@@ -757,9 +757,12 @@ def test_opening_a_recording_with_a_stored_result_runs_no_pass(pass_root):
     controller, host, engine, parts = _controller(host=host)
     assert controller.load_recording(RECORDING) is True
     assert calls == [str(RECORDING)], f"the service was built {len(calls)} times"
-    assert controller.analysis_active(), "the timeline did not adopt the stored result"
-    assert engine.mode() == MODE_ANALYSIS
-    assert parts[2].kinds[-1] == "analysis", "the 2D panel was not told it is a result"
+    # Adoption is "the recording has a readable result", not "the cursor moved onto it":
+    # the engine keeps the recording's own timeline (Part B §9), so the session is held
+    # by the library and answered through ``available_result``.
+    assert controller.available_result() is not None, "the stored result was not adopted"
+    assert engine.mode() == MODE_RECORDING, "adopting a result moved the cursor's timeline"
+    assert parts[2].kinds[-1] == "replay", "the 2D panel was not told the cursor is a replay"
     assert parts[3].kinds[-1] == "analysis", "the 3D pane was not told it is a result"
     assert parts[4].states[-1] == "reused", "the surface did not report the reuse"
     assert stored.is_dir()
@@ -782,8 +785,8 @@ def test_opening_a_recording_without_a_result_starts_the_pass(tmp_path, qt_app):
     assert controller.wait_for_analysis(120)
     _flush()
     assert parts[4].states[-1] == "complete", parts[4].states
-    assert controller.analysis_active(), "a complete result was not adopted"
-    assert engine.mode() == MODE_ANALYSIS
+    assert controller.available_result() is not None, "a complete result was not adopted"
+    assert engine.mode() == MODE_RECORDING, "a finished pass moved the cursor's timeline"
 
 
 def test_a_cancelled_pass_is_reported_and_never_served(tmp_path, qt_app):
diff --git a/apps/studio/tests/test_replay_wiring.py b/apps/studio/tests/test_replay_wiring.py
new file mode 100644
index 0000000..acc0baf
--- /dev/null
+++ b/apps/studio/tests/test_replay_wiring.py
@@ -0,0 +1,507 @@
+"""Part B's wiring: the cached analysis views over a recording, in the running app.
+
+The result contract ([replay-analysis.md](../docs/replay-analysis.md)), the playback
+contract ([replay-analysis-part-b.md](../docs/replay-analysis-part-b.md)) and the panel
+contract ([replay-panel.md](../docs/replay-panel.md)) all describe one state: a recording
+is open, the cursor is on the recording's own signal time, and the arrows and the depth
+map are served from the stored analysis result while the operator scrubs and plays — with
+the recording's own picture one step away in the mode selector, and the recompute switch
+left as the explicit "run the chain over this footage again" action.
+
+What was missing was the wiring that reaches it. ``ReplayEngine.use_recording()`` had no
+caller, so analysis-on-open moved the cursor onto the *result's* timeline instead: the
+derived views were drawn, but the recording's own picture was not offered at all, and the
+cursor could not leave the frames the result covers. These tests drive the window's own
+「打开录像」 action over a recording whose result is already stored beside it — the result
+is generated here into the recording's own cache, so the app *reuses* it and opens no SDK
+chain — and assert on the picture the panel actually shows, not on a state a harness
+built.
+
+The recording is copied before use: a pass writes its result into the recording's own
+directory, and nothing may write into ``dist/``.
+"""
+
+from __future__ import annotations
+
+import shutil
+import time
+from pathlib import Path
+
+import numpy as np
+import pytest
+from orisys_app_shared.acquisition.arrows import depth_map_to_bgr, flow_colormap
+from orisys_app_shared.acquisition.image_backend import (
+    ARROW_PURE,
+    ARROW_STITCHED,
+    DEPTH,
+    DERIVED_MODES,
+    RECORDED,
+)
+from orisys_app_shared.i18n import t
+from orisys_app_shared.scene3d.contracts import DisplaySettings
+from orisys_app_shared.scene3d.visual_policy import (
+    select_flow_samples,
+)
+from orisys_studio.replay import MODE_RECORDING
+from PySide6.QtWidgets import QApplication
+
+REPO = Path(__file__).resolve().parents[3]
+STUDIO = REPO / "dist" / "OrisysStudio"
+SOURCE_RECORDING = (
+    STUDIO
+    / "outputs"
+    / "recordings"
+    / "20260930T015028Z_session0002_40cb8e7441434feab48bb646afdba011"
+)
+#: What makes a directory a recording for the application, and all a replay reads.
+RECORDING_FILES = (
+    "frames.csv",
+    "forces.csv",
+    "events.csv",
+    "session.json",
+    "raw_0001.mp4",
+    "preroll_0001.mp4",
+)
+
+
+def _require_inputs() -> None:
+    missing = [
+        str(SOURCE_RECORDING / name)
+        for name in ("frames.csv", "raw_0001.mp4", "session.json")
+        if not (SOURCE_RECORDING / name).exists()
+    ]
+    if missing:
+        pytest.skip(
+            "the replay wiring tests need the evidence recording: " + ", ".join(missing)
+        )
+    pytest.importorskip("orisys", reason="the analysis pass drives the SDK chain")
+
+
+def _copy_recording(destination: Path) -> Path:
+    """The evidence recording, copied: nothing may write into ``dist/``."""
+    destination.mkdir(parents=True, exist_ok=True)
+    for name in RECORDING_FILES:
+        source = SOURCE_RECORDING / name
+        if source.exists():
+            shutil.copy2(source, destination / name)
+    return destination
+
+
+def _run_the_pass(recording: Path):
+    """One real pass, with the chain the *application* derives from the recording."""
+    from orisys_app_shared.analysis.service import AnalysisService, chain_spec_for
+    from orisys_app_shared.record import RecordingReader
+
+    reader = RecordingReader(recording)
+    try:
+        chain = chain_spec_for(reader.session())
+    finally:
+        reader.close()
+    return AnalysisService(recording, chain=chain).run()
+
+
+@pytest.fixture(scope="module")
+def stored_recording(tmp_path_factory) -> Path:
+    """A copy of the evidence recording whose own cache already holds a result."""
+    _require_inputs()
+    recording = _copy_recording(tmp_path_factory.mktemp("replay-wiring") / "recording")
+    outcome = _run_the_pass(recording)
+    assert outcome.complete, outcome.reason
+    assert outcome.frames_written > 0
+    return recording
+
+
+@pytest.fixture
+def window(make_window, qt_app):
+    """The application, in the replay view, as the operator meets it."""
+    window = make_window(sensor_type="box", source_mode="camera")
+    window.resize(1440, 880)
+    window.show()
+    QApplication.processEvents()
+    window.brand_bar.btn_playback.click()
+    QApplication.processEvents()
+    assert window._in_replay_view(), "the window did not switch to the replay view"
+    return window
+
+
+def _open(window, recording: Path, monkeypatch) -> None:
+    """Press the window's own 「打开录像」 entry, with the file dialog answered for it."""
+    import orisys_studio.replay_controller as replay_controller
+
+    video = Path(recording) / "raw_0001.mp4"
+    monkeypatch.setattr(
+        replay_controller.QFileDialog,
+        "getOpenFileName",
+        lambda *args, **kwargs: (str(video), ""),
+    )
+    action = window.replay_ctl._actions.playback
+    assert action.isEnabled(), "the replay entries are not offered in the replay view"
+    action.trigger()
+    QApplication.processEvents()
+    assert window.replay.has_video(), "the recording did not open"
+
+
+def _sink_frames(window, monkeypatch) -> list:
+    """Every frame handed to the window's single 2D sink, in order."""
+    captured: list = []
+    original = window._show_image_frame
+
+    def spy(frame):
+        captured.append(frame)
+        return original(frame)
+
+    monkeypatch.setattr(window, "_show_image_frame", spy)
+    return captured
+
+
+def _panel_bgr(window) -> np.ndarray:
+    """The 2D panel's picture, in BGR, from the end of the real display path."""
+    image = window.arrow_panel._source_image
+    assert image is not None and not image.isNull(), "the 2D panel shows no picture"
+    width, height = image.width(), image.height()
+    rows = np.frombuffer(bytes(image.constBits()), dtype=np.uint8)
+    rows = rows.reshape(height, image.bytesPerLine())
+    return rows[:, : width * 3].reshape(height, width, 3)[:, :, ::-1].copy()
+
+
+def _expected_depth(window, frame) -> np.ndarray:
+    """A stored frame's depth view, drawn the way the window's own sink draws it."""
+    return depth_map_to_bgr(
+        frame.depth,
+        flow_colormap(window._theme),
+        value_range=window._depth_range(),
+        invalid_mask=window._domain_mask.mask,
+    )
+
+
+def _set_mode(window, mode: str) -> None:
+    index = window.arrow_panel.view_combo.findData(mode)
+    assert index >= 0, f"the mode {mode} is not offered: {window.arrow_panel.view_modes()}"
+    window.arrow_panel.view_combo.setCurrentIndex(index)
+    QApplication.processEvents()
+    assert window.arrow_panel.view_mode() == mode
+
+
+def _pump(seconds: float) -> None:
+    """Let real time pass so the playback timer fires, as it does in the app."""
+    deadline = time.perf_counter() + float(seconds)
+    while time.perf_counter() < deadline:
+        QApplication.processEvents()
+        time.sleep(0.01)
+
+
+def _cells(rows, cols) -> set:
+    return {(int(row), int(col)) for row, col in zip(np.asarray(rows), np.asarray(cols))}
+
+
+def _active_cells(samples) -> set:
+    return {
+        (int(row), int(col))
+        for row, col, active in zip(
+            np.asarray(samples.rows), np.asarray(samples.cols), np.asarray(samples.active)
+        )
+        if bool(active)
+    }
+
+
+def test_the_operator_sees_the_cached_views_over_the_recording(
+    stored_recording, window, monkeypatch
+):
+    """The state the contracts describe, reached the way an operator reaches it.
+
+    A recording with a result beside it opens on the **recording's** timeline (the result
+    is a cache behind the cursor, not a second timeline), the panel offers the recording's
+    own picture *and* the three derived views, and what is on screen is the stored frame —
+    not the camera, and not a placeholder.
+    """
+    captured = _sink_frames(window, monkeypatch)
+    _open(window, stored_recording, monkeypatch)
+
+    assert window.replay.mode() == MODE_RECORDING, (
+        "the cursor left the recording's own timeline; the result is a cache behind it"
+    )
+    session = window.replay_ctl.available_result()
+    assert session is not None, "the stored result beside the recording was not adopted"
+    assert not window.replay_ctl.recompute_active(), (
+        "the cache was only reachable by turning the recompute switch on"
+    )
+    assert window.analysis_progress.state() == "reused", (
+        "a pass ran over a recording that already had a result"
+    )
+
+    assert window.arrow_panel.view_modes() == (
+        RECORDED,
+        ARROW_PURE,
+        ARROW_STITCHED,
+        DEPTH,
+    ), "the recording's own picture must stay one step away from the derived views"
+    assert window.arrow_panel.view_mode() in DERIVED_MODES, (
+        "the operator was left on the camera picture although a result is served"
+    )
+    assert window.arrow_panel.note_label.text() == "", (
+        "the panel says there is nothing to draw while it is drawing a result"
+    )
+
+    # The arrow view is the stored measurement, at the cursor's own frame.
+    frames = list(session.frames())
+    active_frame = next(frame for frame in frames if frame.active_count > 0)
+    window.replay.set_position(session.time_of(active_frame.frame_id))
+    QApplication.processEvents()
+    assert captured, "nothing reached the 2D sink"
+    drawn = captured[-1]
+    assert drawn.frame_id == active_frame.frame_id
+    assert drawn.result_available is True
+    stored_cells = _cells(active_frame.samples.rows, active_frame.samples.cols)
+    field = np.asarray(drawn.flow)
+    finite = np.argwhere(np.isfinite(field[..., 0]))
+    assert _cells(finite[:, 0], finite[:, 1]) == stored_cells, (
+        "the recovered flow field carries values outside the stored candidate locations"
+    )
+    # The height last: the shared selector re-reads the same cells, in its own order, and
+    # comes to the same activity verdict as the pass stored.
+    reselected = select_flow_samples(field, drawn.valid_mask, session.display_policy)
+    assert _cells(reselected.rows, reselected.cols) == stored_cells
+    assert _active_cells(reselected) == _active_cells(active_frame.samples)
+    arrows_bgr = _panel_bgr(window)
+    raw_here = window.replay.video_frame_bgr()
+    assert raw_here is not None
+    assert not np.array_equal(arrows_bgr, np.asarray(raw_here)), (
+        "the arrow view is the recording's raw picture"
+    )
+
+    # The depth view is the stored map, drawn through the window's own backend. This
+    # result's map is exactly zero on every frame it holds — the box chain's depth is
+    # flat on this footage, which the panel has its own sentence for — and it is still a
+    # stored array of the result's own shape, so the pixels are checked against it rather
+    # than against a flat placeholder.
+    _set_mode(window, DEPTH)
+    target = active_frame
+    assert target.depth is not None and target.depth_nonzero == 0
+    QApplication.processEvents()
+    drawn = captured[-1]
+    assert drawn.frame_id == target.frame_id
+    assert np.array_equal(
+        np.asarray(drawn.depth), np.asarray(target.depth)
+    ), "the depth view is not the stored map"
+    shown = _panel_bgr(window)
+    expected = _expected_depth(window, target)
+    assert expected is not None
+    assert shown.shape == expected.shape
+    assert np.array_equal(shown, expected), (
+        "the depth view's pixels are not depth_map_to_bgr of the stored frame"
+    )
+    assert not np.array_equal(shown, arrows_bgr), "the depth view drew the arrow view"
+    assert t("panels.depth_empty", pixels=int(target.depth.size)) in (
+        window.arrow_panel.depth_scale_label.text()
+    ), "an exactly-zero depth map is not stated as such"
+
+    # …and the recording's own picture is still reachable, at the same cursor.
+    _set_mode(window, RECORDED)
+    raw = window.replay.video_frame_bgr()
+    assert raw is not None, "the recording's own frame could not be decoded"
+    assert np.array_equal(_panel_bgr(window), np.asarray(raw)), (
+        "the recorded view is not the recording's own picture"
+    )
+    assert window.arrow_panel.note_label.text() == ""
+    assert window.replay.mode() == MODE_RECORDING
+
+
+def test_playing_forward_and_seeking_draw_the_same_cached_frame(
+    stored_recording, window, monkeypatch
+):
+    """A seek is the frame playing forward passes through — on the operator's path."""
+    captured = _sink_frames(window, monkeypatch)
+    _open(window, stored_recording, monkeypatch)
+    session = window.replay_ctl.available_result()
+    assert session is not None
+    _set_mode(window, DEPTH)
+
+    ids = [int(value) for value in session.frame_ids]
+    target_id = ids[len(ids) // 4]
+    moment = session.time_of(target_id)
+    assert moment is not None
+
+    # Play forward from the top, a tick's worth of cursor at a time, ending on the
+    # target instant itself.
+    positions = [step / 240.0 for step in range(int(moment * 240.0) + 1)]
+    positions.append(moment)
+    for position in positions:
+        window.replay.set_position(position)
+    played = _panel_bgr(window)
+    played_id = captured[-1].frame_id
+
+    # …then seek straight to the same instant.
+    window.replay.set_position(moment)
+    QApplication.processEvents()
+    assert captured[-1].frame_id == played_id == target_id
+    assert np.array_equal(played, _panel_bgr(window)), (
+        "a seek drew a different frame from playing forward to the same position"
+    )
+
+    # Stepping one stored frame on moves to that frame's own picture.
+    later = ids[ids.index(target_id) + 1]
+    window.replay.set_position(session.time_of(later))
+    QApplication.processEvents()
+    assert captured[-1].frame_id == later
+    assert np.array_equal(
+        np.asarray(captured[-1].depth), np.asarray(session.frame(later).depth)
+    )
+
+
+def test_playback_draws_the_cached_frame_not_the_video(
+    stored_recording, window, monkeypatch
+):
+    """Advance the real playback timer and check every drawn frame against the cache."""
+    captured = _sink_frames(window, monkeypatch)
+    _open(window, stored_recording, monkeypatch)
+    session = window.replay_ctl.available_result()
+    assert session is not None
+    _set_mode(window, DEPTH)
+
+    window.replay.set_position(0.0)
+    QApplication.processEvents()
+    captured.clear()
+    window.replay_ctl.start_playback()
+    assert window.replay_ctl.is_playing(), "playback did not start"
+    _pump(0.4)
+    position = window.replay.position()
+    assert position > 0.0, "the playhead did not advance"
+    assert captured, "playback drew nothing"
+    assert position <= session.duration_s(), "the playhead left the recording's timeline"
+
+    expected_id = session.frame_id_at(position)
+    assert expected_id is not None
+    for frame in captured:
+        assert frame.depth is not None, "a frame drawn during playback carries no depth"
+        assert frame.result_available is True
+        assert session.has(int(frame.frame_id)), (
+            f"playback drew frame {frame.frame_id}, which the result does not hold"
+        )
+    # Pausing leaves the frame the cursor is on, and it is the cached one.
+    window.replay_ctl.pause_playback()
+    assert not window.replay_ctl.is_playing()
+    assert captured[-1].frame_id == expected_id
+    expected = _expected_depth(window, session.frame(expected_id))
+    assert expected is not None
+    assert np.array_equal(_panel_bgr(window), expected), (
+        "the paused frame is not the stored frame for the cursor"
+    )
+
+
+def test_a_finished_pass_puts_the_cache_on_screen(tmp_path, window, monkeypatch):
+    """The other trigger: the pass completes and the panel serves what it stored.
+
+    Nothing is pre-placed here, so the window runs the pass over the copy (a box chain:
+    a couple of seconds for this recording) and the *completion* is what has to leave the
+    derived view on screen — with the cursor still on the recording.
+    """
+    recording = _copy_recording(tmp_path / "recording")
+    _sink_frames(window, monkeypatch)
+    _open(window, recording, monkeypatch)
+
+    assert window.replay_ctl.wait_for_analysis(180.0), "the pass did not finish"
+    QApplication.processEvents()
+    assert window.replay_ctl.analysis_error() == "", (
+        f"the pass left no usable result: {window.replay_ctl.analysis_error()}"
+    )
+    assert window.analysis_progress.state() == "complete"
+    session = window.replay_ctl.available_result()
+    assert session is not None, "the completed result was not adopted"
+    assert window.replay.mode() == MODE_RECORDING, (
+        "a finished pass moved the cursor off the recording's own timeline"
+    )
+    assert window.arrow_panel.view_mode() in DERIVED_MODES, (
+        "the operator was left on the camera picture after the pass finished"
+    )
+    assert not window.replay_ctl.recompute_active()
+
+    frame_id = session.frame_id_at(window.replay.position())
+    assert frame_id is not None
+    _set_mode(window, DEPTH)
+    expected = _expected_depth(window, session.frame(frame_id))
+    assert expected is not None
+    assert np.array_equal(_panel_bgr(window), expected), (
+        "the finished result is not what the panel draws"
+    )
+
+
+def test_the_mode_selector_is_the_trigger_for_a_derived_view(
+    stored_recording, window, monkeypatch
+):
+    """On the recording's timeline, choosing a derived view serves it from the cache."""
+    captured = _sink_frames(window, monkeypatch)
+    _open(window, stored_recording, monkeypatch)
+    session = window.replay_ctl.available_result()
+    assert session is not None
+    _set_mode(window, RECORDED)
+
+    position = session.time_of(int(session.frame_ids[len(session.frame_ids) // 2]))
+    assert position is not None
+    window.replay.set_position(position)
+    QApplication.processEvents()
+    assert window.replay.mode() == MODE_RECORDING
+    assert captured[-1].depth is None, "the recorded view drew a stored frame"
+
+    captured.clear()
+    _set_mode(window, DEPTH)
+    assert captured, "choosing a derived view drew nothing"
+    assert captured[-1].depth is not None, (
+        "a view that needs derived data drew the recording's raw picture"
+    )
+    assert captured[-1].frame_id == session.frame_id_at(position)
+    assert window.replay.mode() == MODE_RECORDING, (
+        "choosing a view moved the cursor's timeline"
+    )
+
+
+def test_the_display_settings_still_redraw_from_the_cache(
+    stored_recording, window, monkeypatch
+):
+    """A threshold change is a read: it redraws the stored frame and writes nothing."""
+    from orisys_app_shared.analysis.cache import scan_results
+
+    _open(window, stored_recording, monkeypatch)
+    session = window.replay_ctl.available_result()
+    assert session is not None
+    _set_mode(window, ARROW_PURE)
+
+    cache_root = Path(stored_recording) / "analysis"
+    before = {entry.directory for entry in scan_results(cache_root)}
+    captured = _sink_frames(window, monkeypatch)
+    window.replay_ctl.on_display_settings(DisplaySettings(flow_on_px=0.9))
+    QApplication.processEvents()
+
+    assert session.threshold() == pytest.approx(0.9), (
+        "the display threshold never reached the session"
+    )
+    assert captured, "a display change did not redraw the frame"
+    assert captured[-1].frame_id == session.frame_id_at(window.replay.position())
+    assert {entry.directory for entry in scan_results(cache_root)} == before, (
+        "a display change wrote an analysis result"
+    )
+
+
+def test_a_setting_that_moves_the_grid_is_still_refused(
+    stored_recording, window, monkeypatch
+):
+    """The stored candidate locations cannot be re-gridded; the frame is drawn anyway."""
+    from orisys_app_shared.analysis.cache import scan_results
+
+    _open(window, stored_recording, monkeypatch)
+    session = window.replay_ctl.available_result()
+    assert session is not None
+    _set_mode(window, ARROW_PURE)
+    before_threshold = session.threshold()
+    cache_root = Path(stored_recording) / "analysis"
+    before = {entry.directory for entry in scan_results(cache_root)}
+
+    captured = _sink_frames(window, monkeypatch)
+    window.replay_ctl.on_display_settings(DisplaySettings(max_arrows=50))
+    QApplication.processEvents()
+
+    assert session.threshold() == pytest.approx(before_threshold), (
+        "a refused policy changed the session"
+    )
+    assert captured, "the frame was not redrawn from the result's own grid"
+    assert captured[-1].frame_id == session.frame_id_at(window.replay.position())
+    assert {entry.directory for entry in scan_results(cache_root)} == before