feat(ocr): resolution-anchored layout module + crop preprocessing

Implements ADR-009 (structure + preprocessing; exact IK fixed coordinates still
need calibration against a reference 16:9 screenshot).

- src/lib/layoutProfile.ts: pure, unit-tested geometry for the artifact screen -
  detail rect, the four detail crops, inventory rect/count crop, 5-col grid,
  16:9 detection, aspect label, and an off-16:9 support warning. Single source of
  truth; electron/main.ts now delegates all crop/grid geometry to it and keeps
  colour detection only as the detail-rect fallback.
- src/lib/ocrPreprocess.ts: pure, unit-tested Otsu binarization with inversion
  (artifact text is the bright foreground) over a BGRA bitmap.
- main.ts: OCR now reads an upscaled + binarized copy of each crop; the original
  crop is retained for the diagnostics UI. CaptureResult carries layout info
  { aspect, isSixteenNine, warning }.

NOTE: image preprocessing changes the OCR input and cannot be validated by the
text-level eval harness; it needs a live Genshin 16:9 capture to confirm/tune
(threshold, invert, upscale factor). 88 tests + build green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

README.md

@@ -34,6 +34,20 @@ npm run dev
 
 Use the Electron app window for scanner work. The browser preview does not expose the local capture bridge.
 
+### Input/Capture helper (C# sidecar)
+
+Input automation and screen capture run through a compiled C# sidecar
+(`native/input-helper`, see ADR-008). Build it once:
+
+```powershell
+npm run helper:build   # requires the .NET SDK; produces a self-contained exe
+```
+
+The app auto-detects the exe (`INPUT_HELPER_EXE` env override → packaged
+`resources/input-helper` → `native/input-helper/bin/publish`). If the exe is not
+present it falls back to the embedded PowerShell helper, so the app still runs
+without the .NET build - just slower and with the old per-frame temp-file capture.
+
 ### Automatischer Scan: als Administrator starten
 
 Genshin läuft erhöht (Administrator). Windows (UIPI) verwirft dann alle simulierten Maus-Eingaben aus einer nicht-erhöhten App - SendInput meldet dabei trotzdem Erfolg. Für den automatischen Scan muss die App deshalb ebenfalls erhöht laufen:

electron/main.ts

@@ -1,5 +1,6 @@
 import { app, BrowserWindow, Menu, desktopCapturer, globalShortcut, nativeImage, screen, type NativeImage } from "electron";
 import fs from "node:fs/promises";
+import { existsSync } from "node:fs";
 import http, { type Server } from "node:http";
 import path from "node:path";
 import { fileURLToPath } from "node:url";
@@ -21,6 +22,17 @@ import type {
   ReviewSamplesRepositoryPort,
   ScannerLearningRepositoryPort,
 } from "./repositories/index.js";
+import {
+  aspectRatioLabel,
+  detailCropRects,
+  inventoryCountCropRect,
+  inventoryGrid as layoutInventoryGrid,
+  inventoryRect as layoutInventoryRect,
+  isSixteenNine,
+  layoutSupportWarning,
+  profileDetailRect,
+} from "../src/lib/layoutProfile.js";
+import { binarizeForOcr } from "../src/lib/ocrPreprocess.js";
 
 // Chromium's renderer sandbox can refuse to fully initialize (or silently
 // crash the GPU/renderer process) when the hosting process runs with a full
@@ -70,6 +82,25 @@ function getInputHelperService() {
   return inputHelperService;
 }
 
+// Locate the compiled C# input/capture sidecar (ADR-008). Falls back to null so
+// the service uses the embedded PowerShell helper when the exe was never built.
+function resolveInputHelperExePath(): string | null {
+  const candidates = [
+    process.env.INPUT_HELPER_EXE,
+    path.join(process.resourcesPath, "input-helper", "InputHelper.exe"),
+    path.join(app.getAppPath(), "native", "input-helper", "bin", "publish", "InputHelper.exe"),
+  ].filter((candidate): candidate is string => Boolean(candidate));
+
+  for (const candidate of candidates) {
+    try {
+      if (existsSync(candidate)) return candidate;
+    } catch {
+      // Unreadable path; try the next candidate.
+    }
+  }
+  return null;
+}
+
 function getRepositoryContext() {
   if (!repositoryContext) {
     throw new Error("Repository context has not been initialized.");
@@ -726,84 +757,48 @@ function imageCropDataUrl(sourceImage: NativeImage, rect: Electron.Rectangle, im
   return sourceImage.crop(safeRect).toDataURL();
 }
 
+// Preprocessed copy of a crop for OCR (ADR-009): upscale for more pixels, then
+// grayscale + Otsu-binarize with inversion (artifact text is the bright
+// foreground). The original crop is kept separately for the diagnostics UI.
+function preprocessedCropDataUrl(sourceImage: NativeImage, rect: Electron.Rectangle, imageSize: { width: number; height: number }) {
+  const safeRect = clampCaptureRect(rect, imageSize);
+  const upscaled = sourceImage.crop(safeRect).resize({ width: Math.max(1, safeRect.width * 2), quality: "best" });
+  const size = upscaled.getSize();
+  if (!size.width || !size.height) return upscaled.toDataURL();
+  const binarized = binarizeForOcr({ data: upscaled.getBitmap(), width: size.width, height: size.height });
+  return nativeImage
+    .createFromBitmap(Buffer.from(binarized.data), { width: binarized.width, height: binarized.height })
+    .toDataURL();
+}
+
 function createCrops(
   sourceImage: NativeImage,
   imageSize: { width: number; height: number },
   detailRect: Electron.Rectangle,
   inventoryRect: Electron.Rectangle,
 ) {
-  const templates: CropTemplate[] = [
-    {
-      id: "artifact-title",
-      label: "Artifact title",
-      rect: {
-        x: Math.round(detailRect.x + detailRect.width * 0.055),
-        y: Math.round(detailRect.y + detailRect.height * 0.05),
-        width: Math.round(detailRect.width * 0.82),
-        height: Math.round(detailRect.height * 0.16),
-      },
-    },
-    {
-      id: "artifact-main-stat",
-      label: "Main stat",
-      rect: {
-        x: Math.round(detailRect.x + detailRect.width * 0.055),
-        y: Math.round(detailRect.y + detailRect.height * 0.20),
-        width: Math.round(detailRect.width * 0.82),
-        height: Math.round(detailRect.height * 0.18),
-      },
-    },
-    {
-      id: "artifact-substats",
-      label: "Substats",
-      rect: {
-        x: Math.round(detailRect.x + detailRect.width * 0.055),
-        y: Math.round(detailRect.y + detailRect.height * 0.41),
-        width: Math.round(detailRect.width * 0.82),
-        height: Math.round(detailRect.height * 0.25),
-      },
-    },
-    {
-      id: "artifact-footer",
-      label: "Footer",
-      rect: {
-        x: Math.round(detailRect.x + detailRect.width * 0.055),
-        y: Math.round(detailRect.y + detailRect.height * 0.78),
-        width: Math.round(detailRect.width * 0.82),
-        height: Math.round(detailRect.height * 0.16),
-      },
-    },
-  ];
+  const templates: CropTemplate[] = detailCropRects(detailRect, imageSize);
 
   if (inventoryRect.width > 120 && inventoryRect.height > 80) {
     templates.push({
       id: "inventory-count",
       label: "Inventory count",
-      rect: {
-        x: Math.round(inventoryRect.x + inventoryRect.width * 0.62),
-        y: Math.round(inventoryRect.y + inventoryRect.height * 0.02),
-        width: Math.round(inventoryRect.width * 0.34),
-        height: Math.round(inventoryRect.height * 0.09),
-      },
+      rect: inventoryCountCropRect(inventoryRect, imageSize),
     });
   }
 
   return templates
-    .map((template) => ({
-      ...template,
-      rect: clampCaptureRect(template.rect, imageSize),
-      dataUrl: imageCropDataUrl(sourceImage, template.rect, imageSize),
-    }))
-    .filter((crop) => crop.rect.width > 0 && crop.rect.height > 0)
-    .map((crop) => ({
-      ...crop,
-      rect: {
-        x: crop.rect.x,
-        y: crop.rect.y,
-        width: crop.rect.width,
-        height: crop.rect.height,
-      },
-    }));
+    .map((template) => {
+      const rect = clampCaptureRect(template.rect, imageSize);
+      return {
+        id: template.id,
+        label: template.label,
+        rect,
+        dataUrl: imageCropDataUrl(sourceImage, rect, imageSize),
+        ocrDataUrl: preprocessedCropDataUrl(sourceImage, rect, imageSize),
+      };
+    })
+    .filter((crop) => crop.rect.width > 0 && crop.rect.height > 0);
 }
 
 function inferDetailRect(bitmap: Buffer, imageSize: { width: number; height: number }) {
@@ -845,79 +840,17 @@ function inferDetailRect(bitmap: Buffer, imageSize: { width: number; height: num
     return clampCaptureRect({ x: left, y: top, width: widthGuess, height: heightGuess }, imageSize);
   }
 
-  if (width > 0 && height > 0) {
-    return clampCaptureRect(
-      {
-        x: Math.round(width * 0.50),
-        y: Math.round(height * 0.08),
-        width: Math.round(width * 0.46),
-        height: Math.round(height * 0.74),
-      },
-      imageSize,
-    );
-  }
-
-  return { x: 0, y: 0, width, height };
+  // Colour detection found nothing usable; fall back to the resolution-anchored
+  // profile rect (single source of truth in layoutProfile).
+  return profileDetailRect(imageSize);
 }
 
 function inferInventoryRect(imageSize: { width: number; height: number }, detailRect: Electron.Rectangle) {
-  const { width, height } = imageSize;
-  const preferredWidth = Math.max(140, Math.round(width * 0.48));
-  const x = Math.round(width * 0.03);
-  const y = Math.round(detailRect.y + detailRect.height * 0.09);
-  const availableWidth = Math.max(100, detailRect.x - Math.round(width * 0.04));
-  const panelWidth = Math.max(100, Math.min(preferredWidth, availableWidth));
-  const safeWidth = panelWidth > width * 0.85 ? Math.round(width * 0.55) : panelWidth;
-  return clampCaptureRect(
-    {
-      x,
-      y,
-      width: Math.min(safeWidth, Math.max(width - x - Math.round(width * 0.02), 100)),
-      height: Math.max(140, Math.round(height * 0.70)),
-    },
-    imageSize,
-  );
+  return layoutInventoryRect(imageSize, detailRect);
 }
 
 function inferInventoryGrid(imageSize: { width: number; height: number }, detailRect: Electron.Rectangle) {
-  const inventoryRect = inferInventoryRect(imageSize, detailRect);
-  const cols = 5;
-  if (inventoryRect.width < 160 || inventoryRect.height < 140) {
-    return {
-      centers: [],
-      rows: 0,
-      cols: 0,
-      confidence: 0,
-      source: "missing" as const,
-    };
-  }
-
-  const cellWidth = Math.max(56, Math.round(inventoryRect.width / cols));
-  const stepX = Math.round(cellWidth * 0.96);
-  const stepY = Math.round(cellWidth * 1.03);
-  const visibleRows = Math.max(2, Math.min(6, Math.round(inventoryRect.height / Math.max(stepY, 1))));
-
-  const startX = inventoryRect.x + Math.max(6, Math.round(stepX * 0.45));
-  const startY = inventoryRect.y + Math.max(6, Math.round(stepY * 0.45));
-  const centers = [];
-  for (let row = 0; row < visibleRows; row++) {
-    for (let col = 0; col < cols; col++) {
-      const x = startX + col * stepX;
-      const y = startY + row * stepY;
-      if (x < imageSize.width && y < imageSize.height) {
-        centers.push({ x, y, row, col });
-      }
-    }
-  }
-
-  const trimmed = centers.filter((center) => center.x > 0 && center.y > 0);
-  return {
-    centers: trimmed,
-    rows: visibleRows,
-    cols,
-    confidence: trimmed.length >= cols * 2 ? 76 : trimmed.length >= cols ? 58 : 36,
-    source: "detected" as const,
-  };
+  return layoutInventoryGrid(imageSize, detailRect);
 }
 
 async function buildCaptureResult(
@@ -939,7 +872,9 @@ async function buildCaptureResult(
   const croppedPayload = crops.map((crop) => ({
     id: crop.id,
     label: crop.label,
-    dataUrl: crop.dataUrl,
+    // OCR reads the preprocessed (upscaled + binarized) crop; the original is
+    // kept below for the diagnostics UI.
+    dataUrl: crop.ocrDataUrl ?? crop.dataUrl,
   }));
   const recognized = options.skipOcr ? { ocr: [], timedOut: false } : await runOcrOnCropsWithTimeout(croppedPayload);
 
@@ -970,6 +905,11 @@ async function buildCaptureResult(
     })),
     inventoryGrid: inferInventoryGrid(size, detailRect),
     inventoryCount: count,
+    layout: {
+      aspect: aspectRatioLabel(size),
+      isSixteenNine: isSixteenNine(size),
+      warning: layoutSupportWarning(size),
+    },
   };
 }
 
@@ -1032,7 +972,7 @@ function initializeAppLifecycle() {
     artifactStoreRepository = repositoryContext.artifactStoreRepository;
     reviewSamplesRepository = repositoryContext.reviewSamplesRepository;
     scannerLearningRepository = repositoryContext.scannerLearningRepository;
-    inputHelperService = createInputHelperService({ userDataPath });
+    inputHelperService = createInputHelperService({ userDataPath, exePath: resolveInputHelperExePath() });
 
     registerIpcHandlers({
       focusMainWindow: () => focusMainWindow(),

electron/services/inputHelper.ts

@@ -382,7 +382,7 @@ class InputHelperClient {
   private starting: Promise<void> | null = null;
   private disposed = false;
 
-  constructor(private readonly scriptUserDataPath: string) {}
+  constructor(private readonly options: { scriptUserDataPath: string; exePath?: string | null }) {}
 
   private async ensureStarted() {
     if (this.child) return;
@@ -396,14 +396,31 @@ class InputHelperClient {
   }
 
   private async start() {
-    const scriptPath = path.join(this.scriptUserDataPath, "input-helper.ps1");
+    // Prefer the compiled C# sidecar (ADR-008). If it is missing or fails to
+    // start, fall back to the embedded PowerShell helper so the app keeps working
+    // on machines where the native exe was never built.
+    if (this.options.exePath) {
+      try {
+        await this.startWith(spawn(this.options.exePath, [], { windowsHide: true, stdio: ["pipe", "pipe", "pipe"] }));
+        return;
+      } catch {
+        this.teardownChild();
+      }
+    }
+    await this.startWith(await this.spawnPowershell());
+  }
+
+  private async spawnPowershell() {
+    const scriptPath = path.join(this.options.scriptUserDataPath, "input-helper.ps1");
     await fs.mkdir(path.dirname(scriptPath), { recursive: true });
     await fs.writeFile(scriptPath, INPUT_HELPER_SCRIPT, "utf8");
-
-    const child = spawn("powershell.exe", ["-NoProfile", "-ExecutionPolicy", "Bypass", "-File", scriptPath], {
+    return spawn("powershell.exe", ["-NoProfile", "-ExecutionPolicy", "Bypass", "-File", scriptPath], {
       windowsHide: true,
       stdio: ["pipe", "pipe", "pipe"],
     });
+  }
+
+  private async startWith(child: ChildProcessWithoutNullStreams) {
     child.stdout.setEncoding("utf8");
     child.stdout.on("data", (chunk: string) => this.handleStdout(chunk));
     child.stderr.setEncoding("utf8");
@@ -419,10 +436,22 @@ class InputHelperClient {
     });
     this.child = child;
 
-    // First request compiles the Win32 interop; give it extra time.
+    // The C# sidecar answers ping immediately; the PowerShell fallback compiles
+    // Win32 interop on the first request, so give it extra time.
     await this.send("ping", {}, 20000);
   }
 
+  private teardownChild() {
+    const child = this.child;
+    this.child = null;
+    this.buffer = "";
+    try {
+      child?.kill();
+    } catch {
+      // Child was never spawned or already gone.
+    }
+  }
+
   private handleStdout(chunk: string) {
     this.buffer += chunk;
     let newlineIndex = this.buffer.indexOf("\n");
@@ -485,8 +514,8 @@ export interface InputHelperService {
   dispose(): void;
 }
 
-export function createInputHelperService(options: { userDataPath: string }): InputHelperService {
-  const inputHelper = new InputHelperClient(options.userDataPath);
+export function createInputHelperService(options: { userDataPath: string; exePath?: string | null }): InputHelperService {
+  const inputHelper = new InputHelperClient({ scriptUserDataPath: options.userDataPath, exePath: options.exePath ?? null });
 
   async function request(op: string, params: Record<string, unknown> = {}, timeoutMs = 8000) {
     return inputHelper.request(op, params, timeoutMs);
@@ -590,16 +619,24 @@ export function createInputHelperService(options: { userDataPath: string }): Inp
 
   async function capturePrimaryScreenViaGdi() {
     const result = await request("capture", {}, 15000);
-    const capturePath = String(result.path);
-    const buffer = await fs.readFile(capturePath);
-    await fs.unlink(capturePath).catch(() => undefined);
+    // The C# sidecar returns PNG bytes inline (no temp file). The PowerShell
+    // fallback writes a temp PNG and returns its path.
+    let base64: string;
+    if (typeof result.imageBase64 === "string" && result.imageBase64) {
+      base64 = result.imageBase64;
+    } else {
+      const capturePath = String(result.path);
+      const buffer = await fs.readFile(capturePath);
+      await fs.unlink(capturePath).catch(() => undefined);
+      base64 = buffer.toString("base64");
+    }
     const captureTargetRaw = typeof result.captureTarget === "string" ? result.captureTarget : "";
     const captureTarget: "primary-screen" | "genshin-client" = captureTargetRaw === "primary-screen" || captureTargetRaw === "genshin-client"
       ? captureTargetRaw
       : "primary-screen";
 
     return {
-      dataUrl: `data:image/png;base64,${buffer.toString("base64")}`,
+      dataUrl: `data:image/png;base64,${base64}`,
       width: Number(result.width),
       height: Number(result.height),
       originX: Number(result.originX),

native/input-helper/.gitignore

@@ -0,0 +1,4 @@
+# .NET build outputs — the self-contained exe is built via `npm run helper:build`,
+# not committed (it is ~100 MB).
+bin/
+obj/

native/input-helper/InputHelper.csproj

@@ -0,0 +1,23 @@
+<Project Sdk="Microsoft.NET.Sdk">
+
+  <PropertyGroup>
+    <OutputType>Exe</OutputType>
+    <TargetFramework>net9.0-windows</TargetFramework>
+    <Nullable>enable</Nullable>
+    <ImplicitUsings>enable</ImplicitUsings>
+    <AssemblyName>InputHelper</AssemblyName>
+    <RootNamespace>GenshinAssistant.InputHelper</RootNamespace>
+    <!-- Screen.PrimaryScreen (Forms) + Bitmap/Graphics.CopyFromScreen (Drawing). -->
+    <UseWindowsForms>true</UseWindowsForms>
+    <!-- Single self-contained exe: no .NET install needed on the user's machine. -->
+    <RuntimeIdentifier>win-x64</RuntimeIdentifier>
+    <SelfContained>true</SelfContained>
+    <PublishSingleFile>true</PublishSingleFile>
+    <IncludeNativeLibrariesForSelfExtract>true</IncludeNativeLibrariesForSelfExtract>
+    <InvariantGlobalization>true</InvariantGlobalization>
+    <!-- Per-monitor DPI v2 so SendInput/capture coordinates match a mixed-DPI
+         multi-monitor setup (same reason as the old PowerShell helper). -->
+    <ApplicationManifest>app.manifest</ApplicationManifest>
+  </PropertyGroup>
+
+</Project>

native/input-helper/Program.cs

@@ -0,0 +1,466 @@
+using System.Diagnostics;
+using System.Drawing;
+using System.Drawing.Imaging;
+using System.Runtime.InteropServices;
+using System.Security.Principal;
+using System.Text;
+using System.Text.Json;
+using System.Windows.Forms;
+
+// Long-lived input/capture sidecar for the Genshin Artifact Assistant.
+// Drop-in replacement for the old PowerShell helper (see ADR-008): identical
+// JSON-over-stdin/stdout protocol - one JSON request per line, one JSON response
+// per line - so the Electron-side InputHelperService is unchanged. Win32 interop
+// is compiled once (this is a native exe), and capture returns base64 PNG bytes
+// directly instead of writing a temp file per frame.
+
+namespace GenshinAssistant.InputHelper;
+
+internal static class Program
+{
+    private static IntPtr _genshinHwnd = IntPtr.Zero;
+
+    private static int Main()
+    {
+        // Manifest already declares PerMonitorV2; this is a belt-and-suspenders
+        // call for hosts that ignore the manifest.
+        try { Native.SetProcessDpiAwarenessContext(Native.DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2); }
+        catch { try { Native.SetProcessDpiAwareness(2); } catch { /* oldest fallback */ Native.SetProcessDPIAware(); } }
+
+        Console.OutputEncoding = Encoding.UTF8;
+        var stdout = Console.Out;
+
+        string? line;
+        while ((line = Console.In.ReadLine()) != null)
+        {
+            if (line.Trim().Length == 0) continue;
+
+            var response = new Dictionary<string, object?> { ["id"] = "", ["ok"] = true };
+            try
+            {
+                using var doc = JsonDocument.Parse(line);
+                var root = doc.RootElement;
+                response["id"] = GetString(root, "id");
+                var op = GetString(root, "op");
+                Handle(op, root, response);
+            }
+            catch (Exception ex)
+            {
+                response["ok"] = false;
+                response["error"] = ex.Message;
+            }
+
+            stdout.WriteLine(JsonSerializer.Serialize(response));
+            stdout.Flush();
+        }
+
+        return 0;
+    }
+
+    private static void Handle(string op, JsonElement root, Dictionary<string, object?> response)
+    {
+        switch (op)
+        {
+            case "ping":
+                response["pong"] = true;
+                break;
+
+            case "cursor":
+            {
+                var state = GetCursorState();
+                response["cursorX"] = state.X;
+                response["cursorY"] = state.Y;
+                response["escapePressed"] = state.Escape;
+                response["enterPressed"] = state.Enter;
+                response["f9Pressed"] = state.F9;
+                break;
+            }
+
+            case "runtime":
+            {
+                var hwnd = FindGenshinWindow();
+                var fgHwnd = Native.GetForegroundWindow();
+                response["isElevated"] = IsElevated();
+                response["genshinFound"] = hwnd != IntPtr.Zero;
+                response["genshinHwnd"] = hwnd.ToInt64();
+                response["targetProcess"] = ProcessNameFromHwnd(hwnd);
+                response["foregroundProcess"] = ProcessNameFromHwnd(fgHwnd);
+                response["foregroundHwnd"] = fgHwnd.ToInt64();
+                response["helperPid"] = Environment.ProcessId;
+                break;
+            }
+
+            case "focus":
+            {
+                var info = FocusGenshinWindow();
+                response["focused"] = info.Focused;
+                response["alreadyForeground"] = info.AlreadyForeground;
+                response["foregroundProcess"] = info.ForegroundProcess;
+                response["targetProcess"] = info.TargetProcess;
+                response["genshinFound"] = info.Hwnd != IntPtr.Zero;
+                response["setForegroundResult"] = info.SetForegroundResult;
+                break;
+            }
+
+            case "click":
+            {
+                var info = FocusGenshinWindow();
+                if (info.Focused && !info.AlreadyForeground) Thread.Sleep(120);
+
+                var targetX = GetInt(root, "x");
+                var targetY = GetInt(root, "y");
+                // Bare SetCursorPos then a batched down+up click, matching the
+                // verified Inventory Kamera sequence: no extra move event, no
+                // gap between move and click.
+                Native.SetCursorPos(targetX, targetY);
+                Native.GetCursorPos(out var pt);
+                var onTarget = Math.Abs(targetX - pt.X) <= 2 && Math.Abs(targetY - pt.Y) <= 2;
+                var clickEventsSent = onTarget ? SendMouseClickBatch() : 0u;
+
+                var state = GetCursorState();
+                response["cursorX"] = state.X;
+                response["cursorY"] = state.Y;
+                response["escapePressed"] = state.Escape;
+                response["enterPressed"] = state.Enter;
+                response["f9Pressed"] = state.F9;
+                response["moved"] = onTarget;
+                response["focused"] = info.Focused;
+                response["alreadyForeground"] = info.AlreadyForeground;
+                response["foregroundProcess"] = info.ForegroundProcess;
+                response["targetProcess"] = info.TargetProcess;
+                response["isElevated"] = IsElevated();
+                // Only report a click when the cursor is verifiably on target and
+                // SendInput injected both events; real acceptance is proven later
+                // by the detail-panel fingerprint.
+                response["clicked"] = onTarget && clickEventsSent >= 2;
+                response["inputBlocked"] = onTarget && clickEventsSent < 2;
+                break;
+            }
+
+            case "scroll":
+            {
+                var info = FocusGenshinWindow();
+                if (info.Focused && !info.AlreadyForeground) Thread.Sleep(120);
+
+                if (TryGetInt(root, "x", out var ax) && TryGetInt(root, "y", out var ay))
+                {
+                    Native.SetCursorPos(ax, ay);
+                    Thread.Sleep(30);
+                }
+
+                Native.GetCursorPos(out var pt);
+                response["cursorX"] = pt.X;
+                response["cursorY"] = pt.Y;
+                response["focused"] = info.Focused;
+                response["foregroundProcess"] = info.ForegroundProcess;
+                response["isElevated"] = IsElevated();
+
+                var notches = GetInt(root, "notches");
+                var stepDelta = notches < 0 ? -120 : 120;
+                var count = Math.Min(60, Math.Abs(notches));
+                uint sentTotal = 0;
+                for (var i = 0; i < count; i++)
+                {
+                    sentTotal += SendMouseWheel(stepDelta);
+                    Thread.Sleep(45);
+                }
+                response["notchesSent"] = sentTotal;
+                response["inputBlocked"] = count > 0 && sentTotal == 0;
+                break;
+            }
+
+            case "bounds":
+            {
+                var bounds = GetGenshinClientBounds();
+                if (bounds == null)
+                {
+                    response["found"] = false;
+                }
+                else
+                {
+                    response["found"] = true;
+                    response["left"] = bounds.Value.Left;
+                    response["top"] = bounds.Value.Top;
+                    response["width"] = bounds.Value.Width;
+                    response["height"] = bounds.Value.Height;
+                }
+                break;
+            }
+
+            case "capture":
+            {
+                var bounds = GetGenshinClientBounds();
+                string captureTarget;
+                Rect area;
+                if (bounds == null)
+                {
+                    var screen = Screen.PrimaryScreen!.Bounds;
+                    area = new Rect { Left = screen.Left, Top = screen.Top, Width = screen.Width, Height = screen.Height };
+                    captureTarget = "primary-screen";
+                }
+                else
+                {
+                    area = bounds.Value;
+                    captureTarget = "genshin-client";
+                }
+
+                using var bitmap = new Bitmap(area.Width, area.Height, PixelFormat.Format32bppArgb);
+                using (var graphics = Graphics.FromImage(bitmap))
+                {
+                    graphics.CopyFromScreen(area.Left, area.Top, 0, 0, bitmap.Size, CopyPixelOperation.SourceCopy);
+                }
+
+                using var stream = new MemoryStream();
+                bitmap.Save(stream, ImageFormat.Png);
+                response["imageBase64"] = Convert.ToBase64String(stream.ToArray());
+                response["width"] = area.Width;
+                response["height"] = area.Height;
+                response["originX"] = area.Left;
+                response["originY"] = area.Top;
+                response["captureTarget"] = captureTarget;
+                break;
+            }
+
+            default:
+                response["ok"] = false;
+                response["error"] = "unknown op";
+                break;
+        }
+    }
+
+    private readonly struct Rect
+    {
+        public int Left { get; init; }
+        public int Top { get; init; }
+        public int Width { get; init; }
+        public int Height { get; init; }
+    }
+
+    private struct CursorState
+    {
+        public int X;
+        public int Y;
+        public bool Escape;
+        public bool Enter;
+        public bool F9;
+    }
+
+    private struct FocusInfo
+    {
+        public IntPtr Hwnd;
+        public bool Focused;
+        public bool AlreadyForeground;
+        public string ForegroundProcess;
+        public string TargetProcess;
+        public bool SetForegroundResult;
+    }
+
+    private static CursorState GetCursorState()
+    {
+        Native.GetCursorPos(out var pt);
+        // Only 0x8000 (held right now). The 0x0001 "pressed since last call" bit
+        // is unreliable and fires for ESC presses used to navigate Genshin menus.
+        var esc = (Native.GetAsyncKeyState(0x1B) & 0x8000) != 0;
+        var enter = (Native.GetAsyncKeyState(0x0D) & 0x8000) != 0;
+        var f9 = (Native.GetAsyncKeyState(0x78) & 0x8000) != 0;
+        return new CursorState { X = pt.X, Y = pt.Y, Escape = esc, Enter = enter, F9 = f9 };
+    }
+
+    private static FocusInfo FocusGenshinWindow()
+    {
+        var hwnd = FindGenshinWindow();
+        var info = new FocusInfo
+        {
+            Hwnd = hwnd,
+            ForegroundProcess = "",
+            TargetProcess = ProcessNameFromHwnd(hwnd),
+        };
+        if (hwnd == IntPtr.Zero) return info;
+
+        info.AlreadyForeground = Native.GetForegroundWindow() == hwnd;
+        if (!info.AlreadyForeground)
+        {
+            Native.ShowWindowAsync(hwnd, 9); // SW_RESTORE
+            // No ALT tap: this app and Genshin run at the same (elevated)
+            // integrity level, so SetForegroundWindow succeeds on its own. An ALT
+            // tap would toggle menu-mnemonic mode and swallow the next inputs.
+            info.SetForegroundResult = Native.SetForegroundWindow(hwnd);
+            Thread.Sleep(140);
+        }
+
+        var foreground = Native.GetForegroundWindow();
+        info.Focused = foreground == hwnd;
+        info.ForegroundProcess = ProcessNameFromHwnd(foreground);
+        return info;
+    }
+
+    private static IntPtr FindGenshinWindow()
+    {
+        if (_genshinHwnd != IntPtr.Zero && Native.IsWindow(_genshinHwnd)) return _genshinHwnd;
+
+        foreach (var proc in Process.GetProcesses())
+        {
+            try
+            {
+                var name = proc.ProcessName;
+                if ((name.Contains("GenshinImpact", StringComparison.OrdinalIgnoreCase)
+                     || name.Contains("YuanShen", StringComparison.OrdinalIgnoreCase)
+                     || name.Contains("Genshin", StringComparison.OrdinalIgnoreCase))
+                    && proc.MainWindowHandle != IntPtr.Zero)
+                {
+                    _genshinHwnd = proc.MainWindowHandle;
+                    return _genshinHwnd;
+                }
+            }
+            catch
+            {
+                // Process exited between enumeration and inspection; ignore.
+            }
+        }
+
+        _genshinHwnd = IntPtr.Zero;
+        return _genshinHwnd;
+    }
+
+    private static Rect? GetGenshinClientBounds()
+    {
+        var hwnd = FindGenshinWindow();
+        if (hwnd == IntPtr.Zero) return null;
+        if (!Native.GetClientRect(hwnd, out var rect)) return null;
+
+        var topLeft = new Native.POINT { X = 0, Y = 0 };
+        if (!Native.ClientToScreen(hwnd, ref topLeft)) return null;
+
+        var width = rect.Right - rect.Left;
+        var height = rect.Bottom - rect.Top;
+        if (width <= 0 || height <= 0) return null;
+
+        return new Rect { Left = topLeft.X, Top = topLeft.Y, Width = width, Height = height };
+    }
+
+    private static string ProcessNameFromHwnd(IntPtr hwnd)
+    {
+        if (hwnd == IntPtr.Zero) return "";
+        Native.GetWindowThreadProcessId(hwnd, out var pid);
+        if (pid == 0) return "";
+        try { return Process.GetProcessById((int)pid).ProcessName; }
+        catch { return ""; }
+    }
+
+    private static bool IsElevated()
+    {
+        using var identity = WindowsIdentity.GetCurrent();
+        var principal = new WindowsPrincipal(identity);
+        return principal.IsInRole(WindowsBuiltInRole.Administrator);
+    }
+
+    private static uint SendMouseClickBatch()
+    {
+        var inputs = new Native.INPUT[2];
+        inputs[0].type = 0; // INPUT_MOUSE
+        inputs[0].mi.dwFlags = Native.MOUSEEVENTF_LEFTDOWN;
+        inputs[1].type = 0;
+        inputs[1].mi.dwFlags = Native.MOUSEEVENTF_LEFTUP;
+        return Native.SendInput(2, inputs, Marshal.SizeOf<Native.INPUT>());
+    }
+
+    private static uint SendMouseWheel(int wheelData)
+    {
+        var inputs = new Native.INPUT[1];
+        inputs[0].type = 0;
+        inputs[0].mi.mouseData = unchecked((uint)wheelData);
+        inputs[0].mi.dwFlags = Native.MOUSEEVENTF_WHEEL;
+        return Native.SendInput(1, inputs, Marshal.SizeOf<Native.INPUT>());
+    }
+
+    private static string GetString(JsonElement root, string name)
+        => root.TryGetProperty(name, out var value) ? value.ToString() : "";
+
+    private static int GetInt(JsonElement root, string name)
+        => TryGetInt(root, name, out var value) ? value : 0;
+
+    private static bool TryGetInt(JsonElement root, string name, out int value)
+    {
+        value = 0;
+        if (!root.TryGetProperty(name, out var element)) return false;
+        if (element.ValueKind == JsonValueKind.Number && element.TryGetInt32(out value)) return true;
+        if (element.ValueKind == JsonValueKind.String && int.TryParse(element.GetString(), out value)) return true;
+        return false;
+    }
+}
+
+internal static class Native
+{
+    public const uint MOUSEEVENTF_LEFTDOWN = 0x0002;
+    public const uint MOUSEEVENTF_LEFTUP = 0x0004;
+    public const uint MOUSEEVENTF_WHEEL = 0x0800;
+    public static readonly IntPtr DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2 = new(-4);
+
+    [StructLayout(LayoutKind.Sequential)]
+    public struct POINT { public int X; public int Y; }
+
+    [StructLayout(LayoutKind.Sequential)]
+    public struct RECT { public int Left; public int Top; public int Right; public int Bottom; }
+
+    // Binary-compatible with the Win32 INPUT for mouse-only use on x64:
+    // type(4) + 4 pad + MOUSEINPUT(32) = 40 bytes = sizeof(INPUT).
+    [StructLayout(LayoutKind.Sequential)]
+    public struct MOUSEINPUT
+    {
+        public int dx;
+        public int dy;
+        public uint mouseData;
+        public uint dwFlags;
+        public uint time;
+        public UIntPtr dwExtraInfo;
+    }
+
+    [StructLayout(LayoutKind.Sequential)]
+    public struct INPUT
+    {
+        public int type;
+        public MOUSEINPUT mi;
+    }
+
+    [DllImport("user32.dll")]
+    public static extern bool SetProcessDPIAware();
+
+    [DllImport("shcore.dll")]
+    public static extern int SetProcessDpiAwareness(int value);
+
+    [DllImport("user32.dll")]
+    public static extern bool SetProcessDpiAwarenessContext(IntPtr value);
+
+    [DllImport("user32.dll")]
+    public static extern bool SetCursorPos(int x, int y);
+
+    [DllImport("user32.dll")]
+    public static extern bool GetCursorPos(out POINT lpPoint);
+
+    [DllImport("user32.dll", SetLastError = true)]
+    public static extern bool GetClientRect(IntPtr hWnd, out RECT lpRect);
+
+    [DllImport("user32.dll", SetLastError = true)]
+    public static extern bool ClientToScreen(IntPtr hWnd, ref POINT lpPoint);
+
+    [DllImport("user32.dll")]
+    public static extern short GetAsyncKeyState(int vKey);
+
+    [DllImport("user32.dll", SetLastError = true)]
+    public static extern uint SendInput(uint nInputs, INPUT[] pInputs, int cbSize);
+
+    [DllImport("user32.dll")]
+    public static extern bool SetForegroundWindow(IntPtr hWnd);
+
+    [DllImport("user32.dll")]
+    public static extern bool ShowWindowAsync(IntPtr hWnd, int nCmdShow);
+
+    [DllImport("user32.dll")]
+    public static extern IntPtr GetForegroundWindow();
+
+    [DllImport("user32.dll")]
+    public static extern bool IsWindow(IntPtr hWnd);
+
+    [DllImport("user32.dll")]
+    public static extern uint GetWindowThreadProcessId(IntPtr hWnd, out uint lpdwProcessId);
+}

native/input-helper/app.manifest

@@ -0,0 +1,11 @@
+<?xml version="1.0" encoding="utf-8"?>
+<assembly manifestVersion="1.0" xmlns="urn:schemas-microsoft-com:asm.v1">
+  <assemblyIdentity version="1.0.0.0" name="GenshinAssistant.InputHelper" type="win32" />
+  <application xmlns="urn:schemas-microsoft-com:asm.v3">
+    <windowsSettings>
+      <!-- PerMonitorV2: coordinates stay correct across mixed-DPI monitors. -->
+      <dpiAwareness xmlns="http://schemas.microsoft.com/SMI/2016/WindowsSettings">PerMonitorV2</dpiAwareness>
+      <dpiAware xmlns="http://schemas.microsoft.com/SMI/2005/WindowsSettings">true/pm</dpiAware>
+    </windowsSettings>
+  </application>
+</assembly>

package.json

@@ -15,6 +15,7 @@
     "lint": "tsc --noEmit",
     "test": "vitest run",
     "eval": "vitest run src/eval/ocrEval.test.ts",
+    "helper:build": "dotnet publish native/input-helper/InputHelper.csproj -c Release -o native/input-helper/bin/publish",
     "data:genshin": "node scripts/generate-genshin-data.cjs"
   },
   "dependencies": {
@@ -48,6 +49,15 @@
       "dist-electron/**/*",
       "package.json"
     ],
+    "extraResources": [
+      {
+        "from": "native/input-helper/bin/publish",
+        "to": "input-helper",
+        "filter": [
+          "**/*"
+        ]
+      }
+    ],
     "win": {
       "target": "nsis",
       "requestedExecutionLevel": "requireAdministrator"

src/features/scan/components/hooks/useScanResultCardsModel.ts

@@ -60,19 +60,20 @@ export function useScanResultCardModel({
   parsed,
 }: Pick<ArtifactResultCardProps, "parsed">): ScanResultCardModel {
   const substats = parsed.substats;
+  const rows: Array<[string, ParsedField]> = [
+    ["Name", parsed.fields.name],
+    ["Slot", parsed.fields.slot],
+    ["Level", getLevelField(parsed)],
+    ["Main", mergeField(parsed.fields.mainStat, parsed.fields.mainValue)],
+    ["Set", parsed.fields.setName],
+    ["Equipped", parsed.fields.equipped],
+    ["Substats", parsed.fields.substats],
+  ];
 
   return {
     levelField: getLevelField(parsed),
     quality: resolveQuality(parsed.confidence),
-    fieldRows: [
-      ["Name", parsed.fields.name],
-      ["Slot", parsed.fields.slot],
-      ["Level", getLevelField(parsed)],
-      ["Main", mergeField(parsed.fields.mainStat, parsed.fields.mainValue)],
-      ["Set", parsed.fields.setName],
-      ["Equipped", parsed.fields.equipped],
-      ["Substats", parsed.fields.substats],
-    ].map(([label, field]) => ({
+    fieldRows: rows.map(([label, field]) => ({
       label,
       field,
       confidenceClassName: resolveFieldConfidenceClass(field),

src/features/scan/components/hooks/useScanTopControlsModel.ts

@@ -8,6 +8,7 @@ export interface ScanTopControlsModel {
   canStartAutoScan: boolean;
   canStartManualScan: boolean;
   canCaptureSingle: boolean;
+  autoScanRunning: boolean;
   handleSourceChange: (event: ChangeEvent<HTMLSelectElement>) => void;
   selectGenshinSource: () => void;
   openSettings: () => void;
@@ -41,7 +42,7 @@ export function useScanTopControlsModel({
   bridgeReady,
   isScanning,
   controller,
-}: ScanTopControlsSectionProps): ScanTopControlsModel {
+}: Omit<ScanTopControlsSectionProps, "refreshCaptureSources">): ScanTopControlsModel {
   const {
     setSettingsOpen,
     setDiagnosticsOpen,
@@ -116,6 +117,7 @@ export function useScanTopControlsModel({
     canStartAutoScan: hasSourceSelected && !isScanning && !autoScanRunning && canAutoScan && !requiresAdminForAutoScan,
     canStartManualScan: hasSourceSelected && !isScanning && !autoScanRunning && canCaptureSource,
     canCaptureSingle: hasSourceSelected && !isScanning && !autoScanRunning && canCaptureSource,
+    autoScanRunning,
     handleSourceChange,
     selectGenshinSource,
     openSettings,

src/features/scan/components/modals/hooks/useScanDetailsModalModel.ts

@@ -1,6 +1,6 @@
 import { useCallback, type MouseEvent } from "react";
 import type { ScanDetailsModalProps } from "../types";
-import type { ParsedArtifactCandidate } from "../../../../lib/artifactOcrParser";
+import type { ParsedArtifactCandidate } from "../../../../../lib/artifactOcrParser";
 
 export interface ScanDetailsModalModel {
   closeDetails: () => void;

src/features/scan/components/modals/hooks/useScanDiagnosticsModalModel.ts

@@ -1,5 +1,5 @@
-import { detailFingerprint } from "../../../../lib/autoScanLoop";
-import { sourceVersion } from "../../../../lib/genshinData";
+import { detailFingerprint } from "../../../../../lib/autoScanLoop";
+import { sourceVersion } from "../../../../../lib/genshinData";
 import { useCallback, useMemo, type MouseEvent } from "react";
 import type { ScanDiagnosticsModalProps } from "../types";
 
@@ -39,16 +39,20 @@ export interface ScanDiagnosticsModalModel {
   canSaveReviewSample: boolean;
 }
 
+type ScanDiagnosticsController = ScanDiagnosticsModalProps["controller"];
+
 interface UseScanDiagnosticsModalModelInput extends Pick<
   ScanDiagnosticsModalProps,
   | "setDetailsOpen"
   | "setDiagnosticsOpen"
-  | "saveReviewSample"
-  | "canSaveReviewSample"
   | "latestCapture"
   | "controller"
 > {
   captureStatus: string;
+  // saveReviewSample / canSaveReviewSample live on the controller, not the modal
+  // props; the component wires them through from controller.* .
+  saveReviewSample: ScanDiagnosticsController["saveReviewSample"];
+  canSaveReviewSample: ScanDiagnosticsController["canSaveReviewSample"];
 }
 
 export function useScanDiagnosticsModalModel({

src/features/scan/components/modals/hooks/useScanReviewQueueModalModel.ts

@@ -1,5 +1,5 @@
 import { useCallback, useMemo, type MouseEvent } from "react";
-import type { ReviewSampleAnalysis } from "../../../../lib/reviewSampleAnalysis";
+import type { ReviewSampleAnalysis } from "../../../../../lib/reviewSampleAnalysis";
 import type { ScanReviewQueueModalProps } from "../types";
 
 export interface ScanReviewQueueRow {
@@ -18,8 +18,10 @@ export interface ScanReviewQueueModalModel {
 
 interface UseScanReviewQueueModalModelInput extends Pick<
   ScanReviewQueueModalProps,
-  "setReviewQueueOpen" | "loadReviewQueue"
+  "setReviewQueueOpen"
 > {
+  // loadReviewQueue is wired through from controller.* by the component.
+  loadReviewQueue: ScanReviewQueueModalProps["controller"]["loadReviewQueue"];
   reviewAnalysis: ReviewSampleAnalysis;
   reviewSampleTotal: number;
 }

src/features/scan/components/modals/hooks/useScanSettingsModalModel.ts

@@ -1,6 +1,6 @@
 import { useCallback, type ChangeEvent, type MouseEvent } from "react";
 import type { CaptureResult, RuntimeInfo } from "../../../../../types/global";
-import { clampScanLimit, clampSkipRows } from "../../../../lib/scannerSession";
+import { clampScanLimit, clampSkipRows } from "../../../../../lib/scannerSession";
 
 export interface ScanSettingsModalModel {
   closeSettings: () => void;

src/lib/layoutProfile.test.ts

@@ -0,0 +1,87 @@
+import { describe, expect, it } from "vitest";
+import {
+  aspectRatioLabel,
+  detailCropRects,
+  inventoryCountCropRect,
+  inventoryGrid,
+  inventoryRect,
+  isSixteenNine,
+  layoutSupportWarning,
+  profileDetailRect,
+} from "./layoutProfile";
+
+const HD = { width: 1920, height: 1080 };
+const QHD = { width: 2560, height: 1440 };
+const ULTRAWIDE = { width: 3440, height: 1440 };
+
+describe("layoutProfile", () => {
+  it("detects 16:9 across common resolutions and rejects ultrawide", () => {
+    expect(isSixteenNine(HD)).toBe(true);
+    expect(isSixteenNine(QHD)).toBe(true);
+    expect(isSixteenNine({ width: 3840, height: 2160 })).toBe(true);
+    expect(isSixteenNine(ULTRAWIDE)).toBe(false);
+    expect(isSixteenNine({ width: 1920, height: 1200 })).toBe(false); // 16:10
+  });
+
+  it("labels the aspect ratio", () => {
+    expect(aspectRatioLabel(HD)).toBe("1.78:1");
+    expect(aspectRatioLabel({ width: 0, height: 0 })).toBe("unknown");
+  });
+
+  it("warns only for non-16:9 resolutions", () => {
+    expect(layoutSupportWarning(HD)).toBe("");
+    expect(layoutSupportWarning(QHD)).toBe("");
+    expect(layoutSupportWarning(ULTRAWIDE)).toContain("nicht 16:9");
+    expect(layoutSupportWarning({ width: 0, height: 0 })).toBe("");
+  });
+
+  it("keeps the detail rect inside the image and on the right half", () => {
+    const rect = profileDetailRect(QHD);
+    expect(rect.x).toBeGreaterThanOrEqual(QHD.width * 0.45);
+    expect(rect.x + rect.width).toBeLessThanOrEqual(QHD.width);
+    expect(rect.y + rect.height).toBeLessThanOrEqual(QHD.height);
+  });
+
+  it("produces the four artifact crops in top-to-bottom order, all clamped", () => {
+    const detail = profileDetailRect(QHD);
+    const crops = detailCropRects(detail, QHD);
+    expect(crops.map((crop) => crop.id)).toEqual([
+      "artifact-title",
+      "artifact-main-stat",
+      "artifact-substats",
+      "artifact-footer",
+    ]);
+    let previousY = -1;
+    for (const crop of crops) {
+      expect(crop.rect.x).toBeGreaterThanOrEqual(0);
+      expect(crop.rect.y).toBeGreaterThan(previousY);
+      expect(crop.rect.x + crop.rect.width).toBeLessThanOrEqual(QHD.width);
+      expect(crop.rect.y + crop.rect.height).toBeLessThanOrEqual(QHD.height);
+      previousY = crop.rect.y;
+    }
+  });
+
+  it("places the inventory count crop inside the inventory panel", () => {
+    const detail = profileDetailRect(QHD);
+    const inv = inventoryRect(QHD, detail);
+    const count = inventoryCountCropRect(inv, QHD);
+    expect(count.x).toBeGreaterThanOrEqual(inv.x);
+    expect(count.x + count.width).toBeLessThanOrEqual(QHD.width);
+  });
+
+  it("builds a 5-column inventory grid on the left", () => {
+    const detail = profileDetailRect(QHD);
+    const grid = inventoryGrid(QHD, detail);
+    expect(grid.cols).toBe(5);
+    expect(grid.source).toBe("detected");
+    expect(grid.centers.length).toBeGreaterThanOrEqual(10);
+    expect(grid.centers.every((center) => center.x < detail.x)).toBe(true);
+  });
+
+  it("reports a missing grid when the inventory panel is too small", () => {
+    const tiny = { width: 320, height: 180 };
+    const grid = inventoryGrid(tiny, profileDetailRect(tiny));
+    expect(grid.source).toBe("missing");
+    expect(grid.centers).toHaveLength(0);
+  });
+});

src/lib/layoutProfile.ts

@@ -0,0 +1,205 @@
+// Resolution-anchored layout geometry for the artifact inventory screen
+// (ADR-009). Inventory Kamera's proven approach is to require borderless 16:9 and
+// derive crop/grid coordinates from the client rectangle instead of detecting the
+// panel by colour each frame. This module is the single, pure, unit-tested source
+// of that geometry; electron/main.ts consumes it for cropping and keeps a
+// colour-based detail-rect detector only as a fallback for off-profile setups.
+//
+// NOTE: the per-field detail crop fractions below are the current working values.
+// True IK-style fixed coordinates need calibration against a reference 16:9
+// screenshot; the structure here is what those calibrated numbers slot into.
+
+export interface LayoutRect {
+  x: number;
+  y: number;
+  width: number;
+  height: number;
+}
+
+export interface CropTemplateRect {
+  id: string;
+  label: string;
+  rect: LayoutRect;
+}
+
+export interface InventoryGridLayout {
+  centers: Array<{ x: number; y: number; row: number; col: number }>;
+  rows: number;
+  cols: number;
+  confidence: number;
+  source: "detected" | "missing";
+}
+
+const SIXTEEN_NINE = 16 / 9;
+
+export function aspectRatio(size: { width: number; height: number }): number {
+  if (!size.height) return 0;
+  return size.width / size.height;
+}
+
+export function aspectRatioLabel(size: { width: number; height: number }): string {
+  const ratio = aspectRatio(size);
+  if (ratio === 0) return "unknown";
+  return `${ratio.toFixed(2)}:1`;
+}
+
+// Genshin's UI is authored for 16:9; other aspect ratios letterbox or reflow and
+// the anchored crops no longer line up. Allow a small tolerance for rounding.
+export function isSixteenNine(size: { width: number; height: number }, tolerance = 0.02): boolean {
+  const ratio = aspectRatio(size);
+  if (ratio === 0) return false;
+  return Math.abs(ratio - SIXTEEN_NINE) <= SIXTEEN_NINE * tolerance;
+}
+
+// Empty when the client is a supported 16:9; otherwise a warning explaining that
+// the anchored crops are unreliable off-profile (ADR-009: non-16:9 is explicitly
+// unsupported for the auto scanner).
+export function layoutSupportWarning(size: { width: number; height: number }): string {
+  if (size.width <= 0 || size.height <= 0) return "";
+  if (isSixteenNine(size)) return "";
+  return `Aufloesung ${size.width}x${size.height} ist nicht 16:9 (${aspectRatioLabel(size)}). Der Auto-Scan ist auf 16:9 im randlosen Fenstermodus ausgelegt; die Erkennung kann daneben liegen.`;
+}
+
+export function clampRect(rect: LayoutRect, imageSize: { width: number; height: number }): LayoutRect {
+  const x = Math.max(0, Math.min(imageSize.width - 1, rect.x));
+  const y = Math.max(0, Math.min(imageSize.height - 1, rect.y));
+  const maxWidth = Math.max(1, imageSize.width - x);
+  const maxHeight = Math.max(1, imageSize.height - y);
+  return {
+    x,
+    y,
+    width: Math.max(1, Math.min(maxWidth, rect.width)),
+    height: Math.max(1, Math.min(maxHeight, rect.height)),
+  };
+}
+
+// Anchored guess for the artifact detail panel on the right of the screen. Used
+// as the primary rect for a clean 16:9 client and as the fallback when colour
+// detection cannot find the panel.
+export function profileDetailRect(imageSize: { width: number; height: number }): LayoutRect {
+  const { width, height } = imageSize;
+  if (width <= 0 || height <= 0) return { x: 0, y: 0, width: Math.max(1, width), height: Math.max(1, height) };
+  return clampRect(
+    {
+      x: Math.round(width * 0.5),
+      y: Math.round(height * 0.08),
+      width: Math.round(width * 0.46),
+      height: Math.round(height * 0.74),
+    },
+    imageSize,
+  );
+}
+
+// The four OCR crops inside the detail panel, as fractions of the detail rect.
+export function detailCropRects(detailRect: LayoutRect, imageSize: { width: number; height: number }): CropTemplateRect[] {
+  const templates: CropTemplateRect[] = [
+    {
+      id: "artifact-title",
+      label: "Artifact title",
+      rect: {
+        x: Math.round(detailRect.x + detailRect.width * 0.055),
+        y: Math.round(detailRect.y + detailRect.height * 0.05),
+        width: Math.round(detailRect.width * 0.82),
+        height: Math.round(detailRect.height * 0.16),
+      },
+    },
+    {
+      id: "artifact-main-stat",
+      label: "Main stat",
+      rect: {
+        x: Math.round(detailRect.x + detailRect.width * 0.055),
+        y: Math.round(detailRect.y + detailRect.height * 0.2),
+        width: Math.round(detailRect.width * 0.82),
+        height: Math.round(detailRect.height * 0.18),
+      },
+    },
+    {
+      id: "artifact-substats",
+      label: "Substats",
+      rect: {
+        x: Math.round(detailRect.x + detailRect.width * 0.055),
+        y: Math.round(detailRect.y + detailRect.height * 0.41),
+        width: Math.round(detailRect.width * 0.82),
+        height: Math.round(detailRect.height * 0.25),
+      },
+    },
+    {
+      id: "artifact-footer",
+      label: "Footer",
+      rect: {
+        x: Math.round(detailRect.x + detailRect.width * 0.055),
+        y: Math.round(detailRect.y + detailRect.height * 0.78),
+        width: Math.round(detailRect.width * 0.82),
+        height: Math.round(detailRect.height * 0.16),
+      },
+    },
+  ];
+
+  return templates.map((template) => ({ ...template, rect: clampRect(template.rect, imageSize) }));
+}
+
+export function inventoryCountCropRect(inventoryRect: LayoutRect, imageSize: { width: number; height: number }): LayoutRect {
+  return clampRect(
+    {
+      x: Math.round(inventoryRect.x + inventoryRect.width * 0.62),
+      y: Math.round(inventoryRect.y + inventoryRect.height * 0.02),
+      width: Math.round(inventoryRect.width * 0.34),
+      height: Math.round(inventoryRect.height * 0.09),
+    },
+    imageSize,
+  );
+}
+
+export function inventoryRect(imageSize: { width: number; height: number }, detailRect: LayoutRect): LayoutRect {
+  const { width, height } = imageSize;
+  const preferredWidth = Math.max(140, Math.round(width * 0.48));
+  const x = Math.round(width * 0.03);
+  const y = Math.round(detailRect.y + detailRect.height * 0.09);
+  const availableWidth = Math.max(100, detailRect.x - Math.round(width * 0.04));
+  const panelWidth = Math.max(100, Math.min(preferredWidth, availableWidth));
+  const safeWidth = panelWidth > width * 0.85 ? Math.round(width * 0.55) : panelWidth;
+  return clampRect(
+    {
+      x,
+      y,
+      width: Math.min(safeWidth, Math.max(width - x - Math.round(width * 0.02), 100)),
+      height: Math.max(140, Math.round(height * 0.7)),
+    },
+    imageSize,
+  );
+}
+
+export function inventoryGrid(imageSize: { width: number; height: number }, detailRect: LayoutRect): InventoryGridLayout {
+  const rect = inventoryRect(imageSize, detailRect);
+  const cols = 5;
+  if (rect.width < 160 || rect.height < 140) {
+    return { centers: [], rows: 0, cols: 0, confidence: 0, source: "missing" };
+  }
+
+  const cellWidth = Math.max(56, Math.round(rect.width / cols));
+  const stepX = Math.round(cellWidth * 0.96);
+  const stepY = Math.round(cellWidth * 1.03);
+  const visibleRows = Math.max(2, Math.min(6, Math.round(rect.height / Math.max(stepY, 1))));
+
+  const startX = rect.x + Math.max(6, Math.round(stepX * 0.45));
+  const startY = rect.y + Math.max(6, Math.round(stepY * 0.45));
+  const centers: InventoryGridLayout["centers"] = [];
+  for (let row = 0; row < visibleRows; row++) {
+    for (let col = 0; col < cols; col++) {
+      const x = startX + col * stepX;
+      const y = startY + row * stepY;
+      if (x < imageSize.width && y < imageSize.height) {
+        centers.push({ x, y, row, col });
+      }
+    }
+  }
+
+  const trimmed = centers.filter((center) => center.x > 0 && center.y > 0);
+  return {
+    centers: trimmed,
+    rows: visibleRows,
+    cols,
+    confidence: trimmed.length >= cols * 2 ? 76 : trimmed.length >= cols ? 58 : 36,
+    source: "detected",
+  };
+}

src/lib/ocrPreprocess.test.ts

@@ -0,0 +1,74 @@
+import { describe, expect, it } from "vitest";
+import { binarizeForOcr, computeLuminanceHistogram, otsuThreshold, type Bitmap } from "./ocrPreprocess";
+
+// Build a BGRA bitmap from a grid of [b,g,r] pixels.
+function bitmapFrom(pixels: Array<[number, number, number]>, width: number, height: number): Bitmap {
+  const data = Buffer.alloc(width * height * 4);
+  pixels.forEach(([b, g, r], index) => {
+    data[index * 4] = b;
+    data[index * 4 + 1] = g;
+    data[index * 4 + 2] = r;
+    data[index * 4 + 3] = 255;
+  });
+  return { data, width, height };
+}
+
+describe("ocrPreprocess", () => {
+  it("computes a luminance histogram over all pixels", () => {
+    const bitmap = bitmapFrom([
+      [0, 0, 0],
+      [255, 255, 255],
+      [0, 0, 0],
+      [255, 255, 255],
+    ], 2, 2);
+    const histogram = computeLuminanceHistogram(bitmap);
+    expect(histogram[0]).toBe(2);
+    expect(histogram[255]).toBe(2);
+    expect(histogram.reduce((sum, count) => sum + count, 0)).toBe(4);
+  });
+
+  it("otsu splits a clean bimodal image between the two peaks", () => {
+    const histogram = new Array<number>(256).fill(0);
+    histogram[20] = 50;
+    histogram[220] = 50;
+    const threshold = otsuThreshold(histogram);
+    expect(threshold).toBeGreaterThanOrEqual(20);
+    expect(threshold).toBeLessThan(220);
+  });
+
+  it("otsu is safe on an empty histogram", () => {
+    expect(otsuThreshold(new Array<number>(256).fill(0))).toBe(127);
+  });
+
+  it("inverts bright foreground to black-on-white by default", () => {
+    // Bright text pixel + dark background pixel.
+    const bitmap = bitmapFrom([
+      [255, 255, 255], // bright -> should become black
+      [0, 0, 0], // dark -> should become white
+    ], 2, 1);
+    const out = binarizeForOcr(bitmap, { threshold: 128 });
+    expect([out.data[0], out.data[1], out.data[2]]).toEqual([0, 0, 0]);
+    expect([out.data[4], out.data[5], out.data[6]]).toEqual([255, 255, 255]);
+    expect(out.data[3]).toBe(255);
+  });
+
+  it("keeps bright foreground white when inversion is disabled", () => {
+    const bitmap = bitmapFrom([
+      [255, 255, 255],
+      [0, 0, 0],
+    ], 2, 1);
+    const out = binarizeForOcr(bitmap, { threshold: 128, invertBrightForeground: false });
+    expect(out.data[0]).toBe(255);
+    expect(out.data[4]).toBe(0);
+  });
+
+  it("preserves dimensions and always emits opaque pixels", () => {
+    const bitmap = bitmapFrom(Array.from({ length: 9 }, () => [100, 100, 100] as [number, number, number]), 3, 3);
+    const out = binarizeForOcr(bitmap);
+    expect(out.width).toBe(3);
+    expect(out.height).toBe(3);
+    for (let pixel = 0; pixel < 9; pixel++) {
+      expect(out.data[pixel * 4 + 3]).toBe(255);
+    }
+  });
+});

src/lib/ocrPreprocess.ts

@@ -0,0 +1,101 @@
+// OCR preprocessing for artifact crops (ADR-009). Tesseract reads a clean, high
+// contrast, dark-text-on-light image far more reliably than Genshin's native
+// bright-text-on-dark UI. This binarizes a crop with Otsu thresholding and (by
+// default) inverts, because artifact text is the bright foreground.
+//
+// Works on a raw BGRA bitmap (Electron NativeImage.getBitmap() layout on
+// Windows). Kept pure and channel-order-agnostic for luminance so it is unit
+// testable without Electron. Upscaling is done separately via NativeImage.resize
+// before this runs - interpolated upscaling of small crops is a big Tesseract win
+// and NativeImage does it better than hand-rolled JS.
+
+export interface Bitmap {
+  data: Uint8Array | Buffer;
+  width: number;
+  height: number;
+}
+
+export interface BinarizeOptions {
+  /** Artifact text is the bright foreground, so invert to dark-on-light. */
+  invertBrightForeground?: boolean;
+  /** Override Otsu with a fixed 0-255 luminance threshold. */
+  threshold?: number;
+}
+
+const BYTES_PER_PIXEL = 4;
+
+// Rec. 601 luma. Channel order does not matter for a weighted sum as long as we
+// read the same three bytes; BGRA and RGBA give the same luminance here because
+// we weight by position-independent coefficients applied to the actual R/G/B.
+function luminanceAt(data: Uint8Array | Buffer, index: number): number {
+  // NativeImage on Windows is BGRA: byte0=B, byte1=G, byte2=R.
+  const b = data[index];
+  const g = data[index + 1];
+  const r = data[index + 2];
+  return 0.299 * r + 0.587 * g + 0.114 * b;
+}
+
+export function computeLuminanceHistogram(bitmap: Bitmap): number[] {
+  const histogram = new Array<number>(256).fill(0);
+  const { data, width, height } = bitmap;
+  const pixels = width * height;
+  for (let pixel = 0; pixel < pixels; pixel++) {
+    const value = Math.round(luminanceAt(data, pixel * BYTES_PER_PIXEL));
+    histogram[Math.max(0, Math.min(255, value))]++;
+  }
+  return histogram;
+}
+
+// Otsu's method: pick the threshold that maximizes between-class variance.
+export function otsuThreshold(histogram: readonly number[]): number {
+  const total = histogram.reduce((sum, count) => sum + count, 0);
+  if (total === 0) return 127;
+
+  let sumAll = 0;
+  for (let level = 0; level < 256; level++) sumAll += level * histogram[level];
+
+  let sumBackground = 0;
+  let weightBackground = 0;
+  let maxVariance = -1;
+  let threshold = 127;
+
+  for (let level = 0; level < 256; level++) {
+    weightBackground += histogram[level];
+    if (weightBackground === 0) continue;
+    const weightForeground = total - weightBackground;
+    if (weightForeground === 0) break;
+
+    sumBackground += level * histogram[level];
+    const meanBackground = sumBackground / weightBackground;
+    const meanForeground = (sumAll - sumBackground) / weightForeground;
+    const betweenVariance = weightBackground * weightForeground * (meanBackground - meanForeground) ** 2;
+
+    if (betweenVariance > maxVariance) {
+      maxVariance = betweenVariance;
+      threshold = level;
+    }
+  }
+
+  return threshold;
+}
+
+export function binarizeForOcr(bitmap: Bitmap, options: BinarizeOptions = {}): Bitmap {
+  const { data, width, height } = bitmap;
+  const invert = options.invertBrightForeground ?? true;
+  const threshold = options.threshold ?? otsuThreshold(computeLuminanceHistogram(bitmap));
+
+  const output = Buffer.alloc(width * height * BYTES_PER_PIXEL);
+  const pixels = width * height;
+  for (let pixel = 0; pixel < pixels; pixel++) {
+    const index = pixel * BYTES_PER_PIXEL;
+    const isBright = luminanceAt(data, index) > threshold;
+    // Bright foreground text -> black; dark background -> white (inverted).
+    const value = invert ? (isBright ? 0 : 255) : (isBright ? 255 : 0);
+    output[index] = value;
+    output[index + 1] = value;
+    output[index + 2] = value;
+    output[index + 3] = 255;
+  }
+
+  return { data: output, width, height };
+}

src/types/global.d.ts

@@ -56,6 +56,11 @@ export interface CaptureResult {
     source: "ocr" | "missing";
     text: string;
   };
+  layout?: {
+    aspect: string;
+    isSixteenNine: boolean;
+    warning: string;
+  };
 }
 
 export interface WindowBounds {