@mytec: iter3.7.0 start, gpu calc int

2026-02-03 22:41:08 +02:00
parent a61753c642
commit 6cd9d869cc
29 changed files with 2288 additions and 28 deletions
--- a/.claude/settings.local.json
+++ b/.claude/settings.local.json
@@ -45,7 +45,8 @@
      "Bash(journalctl:*)",
      "Bash(pkill:*)",
      "Bash(pip3 list:*)",
-      "Bash(chmod:*)"
+      "Bash(chmod:*)",
+      "Bash(pyinstaller:*)"
    ]
  }
 }
--- a/.gitignore
+++ b/.gitignore
@@ -24,3 +24,7 @@ installer/dist/
 __pycache__/
 *.pyc
 nul
+
+# PyInstaller build artifacts
+backend/build/
+backend/dist/
--- a/RFCP-3.6.0-GPU-Build-Task.md
+++ b/RFCP-3.6.0-GPU-Build-Task.md
@@ -0,0 +1,130 @@
+# RFCP 3.6.0 — Production GPU Build (Claude Code Task)
+
+## Goal
+
+Build `rfcp-server.exe` (PyInstaller) with CuPy GPU support so production RFCP
+detects the NVIDIA GPU without manual `pip install`.
+
+Currently production exe shows "CPU (NumPy)" because CuPy is not bundled.
+
+## Current Environment (CONFIRMED WORKING)
+
+```
+Windows 10 (10.0.26200)
+Python 3.11.8 (C:\Python311)
+NVIDIA GeForce RTX 4060 Laptop GPU (8 GB VRAM)
+CUDA Toolkit 13.1 (C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1)
+CUDA_PATH = C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1
+
+Packages:
+  cupy-cuda13x  13.6.0   ← NOT cuda12x!
+  numpy         1.26.4
+  scipy         1.17.0
+  fastrlock     0.8.3
+  pyinstaller   6.18.0
+
+GPU compute verified:
+  python -c "import cupy; a = cupy.array([1,2,3]); print(a.sum())"  → 6 ✅
+```
+
+## What We Already Tried (And Why It Failed)
+
+### Attempt 1: ONEFILE spec with collect_all('cupy')
+- `collect_all('cupy')` returns 1882 datas, **0 binaries** — CuPy pip doesn't bundle DLLs on Windows
+- CUDA DLLs come from two separate sources:
+  - **nvidia pip packages** (14 DLLs in `C:\Python311\Lib\site-packages\nvidia\*/bin/`)
+  - **CUDA Toolkit** (13 DLLs in `C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin\x64\`)
+- We manually collected these 27 DLLs in the spec
+- Build succeeded (3 GB exe!) but crashed on launch:
+  ```
+  [PYI-10456:ERROR] Failed to extract cufft64_12.dll: decompression resulted in return code -1!
+  ```
+- Root cause: `cufft64_12.dll` is 297 MB — PyInstaller's zlib compression fails on it in ONEFILE mode
+
+### Attempt 2: We were about to try ONEDIR but haven't built it yet
+
+### Key Insight: Duplicate DLLs from two sources
+nvidia pip packages have CUDA 12.x DLLs (cublas64_12.dll etc.)
+CUDA Toolkit 13.1 has CUDA 13.x DLLs (cublas64_13.dll etc.)
+CuPy-cuda13x needs the 13.x versions. The 12.x from pip may conflict.
+
+## What Needs To Happen
+
+1. **Build rfcp-server as ONEDIR** (folder with exe + DLLs, not single exe)
+   - This avoids the decompression crash with large CUDA DLLs
+   - Output: `backend/dist/rfcp-server/rfcp-server.exe` + all DLLs alongside
+
+2. **Include ONLY the correct CUDA DLLs** 
+   - Prefer CUDA Toolkit 13.1 DLLs (match cupy-cuda13x)
+   - The nvidia pip packages have cuda12x DLLs — may cause version conflicts
+   - Key DLLs needed: cublas, cusparse, cusolver, curand, cufft, nvrtc, cudart
+
+3. **Exclude bloat** — the previous build pulled in tensorflow, grpc, opentelemetry etc.
+   making it 3 GB. Real size should be ~600-800 MB.
+
+4. **Test the built exe** — run it standalone and verify:
+   - `curl http://localhost:8090/api/health` returns `"build": "gpu"`
+   - `curl http://localhost:8090/api/gpu/status` returns `"available": true`
+   - Or at minimum: the exe starts without errors and CuPy imports successfully
+
+5. **Update Electron integration** if needed:
+   - Current Electron expects a single `rfcp-server.exe` file
+   - With ONEDIR, it's a folder `rfcp-server/rfcp-server.exe`
+   - File: `desktop/main.js` or `desktop/src/main.ts` — look for where it spawns backend
+   - The path needs to change from `resources/backend/rfcp-server.exe` 
+     to `resources/backend/rfcp-server/rfcp-server.exe`
+
+## File Locations
+
+```
+D:\root\rfcp\
+├── backend\
+│   ├── run_server.py          ← PyInstaller entry point
+│   ├── app\
+│   │   ├── main.py            ← FastAPI app
+│   │   ├── services\
+│   │   │   ├── gpu_backend.py ← GPU detection (CuPy/NumPy fallback)
+│   │   │   └── coverage_service.py ← Uses get_array_module()
+│   │   └── api\routes\gpu.py  ← /api/gpu/status, /api/gpu/diagnostics
+│   ├── dist\                  ← PyInstaller output goes here
+│   └── build\                 ← PyInstaller build cache
+├── installer\
+│   ├── rfcp-server-gpu.spec   ← GPU spec (needs fixing)
+│   ├── rfcp-server.spec       ← CPU spec (working, don't touch)
+│   ├── rfcp.ico               ← Icon (exists)
+│   └── build-gpu.bat          ← Build script
+├── desktop\
+│   ├── main.js or src/main.ts ← Electron main process
+│   └── resources\backend\     ← Where production exe lives
+└── frontend\                  ← React frontend (no changes needed)
+```
+
+## Existing CPU spec for reference
+
+The working CPU-only spec is at `installer/rfcp-server.spec`. Use it as the base
+and ADD CuPy + CUDA on top. Don't reinvent the wheel.
+
+## Build Command
+
+```powershell
+cd D:\root\rfcp\backend
+pyinstaller ..\installer\rfcp-server-gpu.spec --clean --noconfirm
+```
+
+## Success Criteria
+
+- [ ] `dist/rfcp-server/rfcp-server.exe` starts without errors
+- [ ] CuPy imports successfully inside the exe (no missing DLL errors)
+- [ ] `/api/gpu/status` returns `"available": true, "device": "RTX 4060"`
+- [ ] Total folder size < 1 GB (ideally 600-800 MB)
+- [ ] No tensorflow/grpc/opentelemetry bloat
+- [ ] Electron can find and launch the backend (path updated if needed)
+
+## Important Notes
+
+- Do NOT use cupy-cuda12x — we migrated to cupy-cuda13x
+- Do NOT try ONEFILE mode — cufft64_12.dll (297 MB) crashes decompression
+- The nvidia pip packages (nvidia-cublas-cu12, etc.) are still installed but may
+  conflict with CUDA Toolkit 13.1 — prefer Toolkit DLLs
+- `collect_all('cupy')` gives 0 binaries on Windows — DLLs must be manually specified
+- gpu_backend.py already handles CuPy absence gracefully (falls back to NumPy)
--- a/RFCP-3.7.0-GPU-Coverage-Task.md
+++ b/RFCP-3.7.0-GPU-Coverage-Task.md
@@ -0,0 +1,133 @@
+# RFCP 3.7.0 — GPU-Accelerated Coverage Calculations
+
+## Context
+
+Iteration 3.6.0 completed: CuPy-cuda13x works in production PyInstaller build,
+RTX 4060 detected, ONEDIR build with CUDA DLLs. BUT coverage calculations still
+run on CPU because coverage_service.py uses `import numpy as np` directly instead
+of the GPU backend.
+
+The GPU infrastructure is ready:
+- `app/services/gpu_backend.py` has `GPUManager.get_array_module()` → returns cupy or numpy
+- `/api/gpu/status` confirms `"active_backend": "cuda"` 
+- CuPy is imported and GPU detected in the frozen exe
+
+## Goal
+
+Replace direct `np.` calls in coverage_service.py with `xp = gpu_manager.get_array_module()`
+so calculations run on GPU when available, with automatic NumPy fallback.
+
+## Files to Modify
+
+### `app/services/coverage_service.py`
+
+**Line 7**: `import numpy as np` — keep this but also import gpu_manager
+
+Add near top:
+```python
+from app.services.gpu_backend import gpu_manager
+```
+
+**Key sections to GPU-accelerate** (highest impact first):
+
+#### 1. Grid array creation (lines 549-550, 922-923)
+```python
+# BEFORE:
+grid_lats = np.array([lat for lat, lon in grid])
+grid_lons = np.array([lon for lat, lon in grid])
+
+# AFTER:
+xp = gpu_manager.get_array_module()
+grid_lats = xp.array([lat for lat, lon in grid])
+grid_lons = xp.array([lon for lat, lon in grid])
+```
+
+#### 2. Trig calculations (line 468, 1031, 1408-1415, 1442)
+These use np.cos, np.radians, np.sin, np.degrees, np.arctan2 — all have CuPy equivalents.
+```python
+# BEFORE:
+lon_delta = settings.radius / (111000 * np.cos(np.radians(center_lat)))
+cos_lat = np.cos(np.radians(center_lat))
+
+# AFTER:
+xp = gpu_manager.get_array_module()
+lon_delta = settings.radius / (111000 * float(xp.cos(xp.radians(center_lat))))
+cos_lat = float(xp.cos(xp.radians(center_lat)))
+```
+
+#### 3. The heavy calculation loop — `_run_point_loop` (line 1070) and `_calculate_point_sync` (line 1112)
+This is where 90% of time is spent. Currently processes points one-by-one.
+The GPU win comes from vectorizing the path loss calculation across ALL grid points at once.
+
+**Strategy**: Instead of looping through points, create arrays of all distances/angles
+and compute path loss for all points in one vectorized operation.
+
+#### 4. `_calculate_bearing` (line 1402) — already vectorizable
+```python
+# All np.* functions here have direct CuPy equivalents
+# Just replace np → xp
+```
+
+## Important Rules
+
+1. **Always get xp at function scope**, not module scope:
+   ```python
+   def my_function(self, ...):
+       xp = gpu_manager.get_array_module()
+       # use xp instead of np
+   ```
+
+2. **Convert GPU arrays back to CPU** before returning to non-GPU code:
+   ```python
+   if hasattr(result, 'get'):  # CuPy array
+       result = result.get()  # → numpy array
+   ```
+
+3. **Keep np for small/scalar operations** — GPU overhead isn't worth it for single values.
+   Only use xp for array operations on 100+ elements.
+
+4. **Don't break the fallback** — if CuPy isn't available, `get_array_module()` returns numpy,
+   so `xp.array()` etc. work identically.
+
+5. **Test both paths** — run with GPU and verify same results as CPU.
+
+## Testing
+
+After changes:
+```powershell
+# Rebuild
+cd D:\root\rfcp\backend
+pyinstaller ..\installer\rfcp-server-gpu.spec --noconfirm
+
+# Run
+.\dist\rfcp-server\rfcp-server.exe
+
+# Test calculation via frontend — watch Task Manager GPU utilization
+# Should see GPU Compute spike during coverage calculation
+# Time should be significantly faster than 10s for 1254 points
+```
+
+Compare before/after:
+- Current (CPU): ~10s for 1254 points, 5km radius
+- Expected (GPU): 1-3s for same calculation
+
+Also test GPU diagnostics:
+```
+curl http://localhost:8888/api/gpu/diagnostics
+```
+
+## What NOT to Change
+
+- Don't modify gpu_backend.py — it's working correctly
+- Don't change the API endpoints or response format
+- Don't remove the NumPy import — keep it for non-array operations
+- Don't change propagation model math — only the array operations
+- Don't change _filter_buildings_to_bbox or OSM functions — they use lists not arrays
+
+## Success Criteria
+
+- [ ] Coverage calculation uses GPU (visible in Task Manager)
+- [ ] Calculation time reduced for 1000+ point grids
+- [ ] CPU fallback still works (test by setting active_backend to cpu via API)
+- [ ] Same coverage results (heatmap should look identical)
+- [ ] No regression in tiled processing mode
--- a/RFCP-Iteration-3.6.0-Production-GPU-Build.md
+++ b/RFCP-Iteration-3.6.0-Production-GPU-Build.md
@@ -0,0 +1,556 @@
+# RFCP — Iteration 3.6.0: Production GPU Build
+
+## Overview
+
+Enable GPU acceleration in the production PyInstaller build. Currently production
+runs CPU-only (NumPy) because CuPy is not included in rfcp-server.exe.
+
+**Goal:** User with NVIDIA GPU installs RFCP → GPU detected automatically → 
+coverage calculations use CUDA acceleration. No manual pip install required.
+
+**Context from diagnostics screenshot:**
+```json
+{
+  "python_executable": "C:\\Users\\Administrator\\AppData\\Local\\Programs\\RFCP\\resources\\backend\\rfcp-server.exe",
+  "platform": "Windows-10-10.0.26288-SP0",
+  "is_wsl": false,
+  "numpy": { "version": "1.26.4" },
+  "cuda": {
+    "error": "CuPy not installed",
+    "install_hint": "pip install cupy-cuda12x"
+  }
+}
+```
+
+**Architecture:** Production uses PyInstaller-bundled rfcp-server.exe (self-contained).
+CuPy not included → GPU not available for end users.
+
+---
+
+## Strategy: Two-Tier Build
+
+Instead of one massive binary, produce two builds:
+
+```
+RFCP-Setup-{version}.exe          (~150 MB) — CPU-only, works everywhere
+RFCP-Setup-{version}-GPU.exe      (~700 MB) — includes CuPy + CUDA runtime
+```
+
+**Why not dynamic loading?**
+PyInstaller bundles everything at build time. CuPy can't be pip-installed 
+into a frozen exe at runtime. Options are:
+
+1. **Bundle CuPy in PyInstaller** ← cleanest, what we'll do
+2. Side-load CuPy DLLs (fragile, version-sensitive)  
+3. Hybrid: unfrozen Python + CuPy installed separately (defeats purpose of exe)
+
+---
+
+## Task 1: PyInstaller Spec with CuPy (Priority 1 — 30 min)
+
+### File: `installer/rfcp-server-gpu.spec`
+
+Create a separate .spec file that includes CuPy:
+
+```python
+# rfcp-server-gpu.spec — GPU-enabled build
+import os
+import sys
+from PyInstaller.utils.hooks import collect_all, collect_dynamic_libs
+
+backend_path = os.path.abspath(os.path.join(os.path.dirname(SPEC), '..', 'backend'))
+
+# Collect CuPy and its CUDA dependencies
+cupy_datas, cupy_binaries, cupy_hiddenimports = collect_all('cupy')
+# Also collect cupy_backends
+cupyb_datas, cupyb_binaries, cupyb_hiddenimports = collect_all('cupy_backends')
+
+# CUDA runtime libraries that CuPy needs
+cuda_binaries = collect_dynamic_libs('cupy')
+
+a = Analysis(
+    [os.path.join(backend_path, 'run_server.py')],
+    pathex=[backend_path],
+    binaries=cupy_binaries + cupyb_binaries + cuda_binaries,
+    datas=[
+        (os.path.join(backend_path, 'data', 'terrain'), 'data/terrain'),
+    ] + cupy_datas + cupyb_datas,
+    hiddenimports=[
+        # Existing imports from rfcp-server.spec
+        'uvicorn.logging',
+        'uvicorn.loops',
+        'uvicorn.loops.auto',
+        'uvicorn.protocols',
+        'uvicorn.protocols.http',
+        'uvicorn.protocols.http.auto',
+        'uvicorn.protocols.websockets',
+        'uvicorn.protocols.websockets.auto',
+        'uvicorn.lifespan',
+        'uvicorn.lifespan.on',
+        'motor',
+        'pymongo',
+        'numpy',
+        'scipy',
+        'shapely',
+        'shapely.geometry',
+        'shapely.ops',
+        # CuPy-specific
+        'cupy',
+        'cupy.cuda',
+        'cupy.cuda.runtime',
+        'cupy.cuda.driver',
+        'cupy.cuda.memory',
+        'cupy.cuda.stream',
+        'cupy._core',
+        'cupy._core.core',
+        'cupy._core._routines_math',
+        'cupy.fft',
+        'cupy.linalg',
+        'fastrlock',
+    ] + cupy_hiddenimports + cupyb_hiddenimports,
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[],
+    excludes=[],
+    noarchive=False,
+)
+
+pyz = PYZ(a.pure)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    a.binaries,
+    a.datas,
+    [],
+    name='rfcp-server',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=False,  # Don't compress CUDA libs — they need fast loading
+    console=True,
+    icon=os.path.join(os.path.dirname(SPEC), 'rfcp.ico'),
+)
+```
+
+### Key Points:
+- `collect_all('cupy')` grabs all CuPy submodules + CUDA DLLs
+- `fastrlock` is a CuPy dependency (must be in hiddenimports)
+- `upx=False` — don't compress CUDA binaries (breaks them)
+- One-file mode (`a.binaries + a.datas` in EXE) for single exe
+
+---
+
+## Task 2: Build Script for GPU Variant (Priority 1 — 15 min)
+
+### File: `installer/build-gpu.bat` (Windows)
+
+```batch
+@echo off
+echo ========================================
+echo  RFCP GPU Build — rfcp-server-gpu.exe
+echo ========================================
+
+REM Ensure CuPy is installed in build environment
+echo Checking CuPy installation...
+python -c "import cupy; print(f'CuPy {cupy.__version__} with CUDA {cupy.cuda.runtime.runtimeGetVersion()}')"
+if errorlevel 1 (
+    echo ERROR: CuPy not installed. Run: pip install cupy-cuda12x
+    exit /b 1
+)
+
+REM Build with GPU spec
+echo Building rfcp-server with GPU support...
+cd /d %~dp0\..\backend
+pyinstaller ..\installer\rfcp-server-gpu.spec --clean --noconfirm
+
+echo.
+echo Build complete! Output: dist\rfcp-server.exe
+echo Size:
+dir dist\rfcp-server.exe
+
+REM Optional: copy to Electron resources
+if exist "..\desktop\resources" (
+    copy /y dist\rfcp-server.exe ..\desktop\resources\rfcp-server.exe
+    echo Copied to desktop\resources\
+)
+
+pause
+```
+
+### File: `installer/build-gpu.sh` (WSL/Linux)
+
+```bash
+#!/bin/bash
+set -e
+
+echo "========================================"
+echo " RFCP GPU Build — rfcp-server (GPU)"
+echo "========================================"
+
+# Check CuPy
+python3 -c "import cupy; print(f'CuPy {cupy.__version__}')" 2>/dev/null || {
+    echo "ERROR: CuPy not installed. Run: pip install cupy-cuda12x"
+    exit 1
+}
+
+cd "$(dirname "$0")/../backend"
+pyinstaller ../installer/rfcp-server-gpu.spec --clean --noconfirm
+
+echo ""
+echo "Build complete!"
+ls -lh dist/rfcp-server*
+```
+
+---
+
+## Task 3: GPU Backend — Graceful CuPy Detection (Priority 1 — 15 min)
+
+### File: `backend/app/services/gpu_backend.py`
+
+The existing gpu_backend.py should already handle CuPy absence gracefully.
+Verify and fix if needed:
+
+```python
+# gpu_backend.py — must work in BOTH CPU and GPU builds
+
+import numpy as np
+
+# Try importing CuPy — this is the key detection
+_cupy_available = False
+_gpu_device_name = None
+_gpu_memory_mb = 0
+
+try:
+    import cupy as cp
+    # Verify we can actually use it (not just import)
+    device = cp.cuda.Device(0)
+    _gpu_device_name = device.attributes.get('name', f'CUDA Device {device.id}')
+    # Try to get name via runtime
+    try:
+        props = cp.cuda.runtime.getDeviceProperties(0)
+        _gpu_device_name = props.get('name', _gpu_device_name)
+        if isinstance(_gpu_device_name, bytes):
+            _gpu_device_name = _gpu_device_name.decode('utf-8').strip('\x00')
+    except Exception:
+        pass
+    _gpu_memory_mb = device.mem_info[1] // (1024 * 1024)
+    _cupy_available = True
+except ImportError:
+    cp = None  # CuPy not installed (CPU build)
+except Exception as e:
+    cp = None  # CuPy installed but CUDA not available
+    print(f"[GPU] CuPy found but CUDA unavailable: {e}")
+
+
+def is_gpu_available() -> bool:
+    return _cupy_available
+
+def get_gpu_info() -> dict:
+    if _cupy_available:
+        return {
+            "available": True,
+            "backend": "CuPy (CUDA)",
+            "device": _gpu_device_name,
+            "memory_mb": _gpu_memory_mb,
+        }
+    return {
+        "available": False,
+        "backend": "NumPy (CPU)",
+        "device": "CPU",
+        "memory_mb": 0,
+    }
+
+def get_array_module():
+    """Return cupy if available, otherwise numpy."""
+    if _cupy_available:
+        return cp
+    return np
+```
+
+### Usage in coverage_service.py:
+
+```python
+from app.services.gpu_backend import get_array_module, is_gpu_available
+
+xp = get_array_module()  # cupy or numpy — same API
+
+# All calculations use xp instead of np:
+distances = xp.sqrt(dx**2 + dy**2)
+path_loss = 20 * xp.log10(distances) + 20 * xp.log10(freq_mhz) - 27.55
+
+# If using cupy, results need to come back to CPU for JSON serialization:
+if is_gpu_available():
+    results = xp.asnumpy(path_loss)
+else:
+    results = path_loss
+```
+
+---
+
+## Task 4: GPU Status in Frontend Header (Priority 2 — 10 min)
+
+### Update GPUIndicator.tsx
+
+When GPU is detected, the badge should clearly show it:
+
+```
+CPU build:     [⚙ CPU]          (gray badge)
+GPU detected:  [⚡ RTX 4060]     (green badge)
+```
+
+The existing GPUIndicator already does this. Just verify:
+1. Badge color changes from gray → green when GPU available
+2. Dropdown shows "Active: GPU (CUDA)" not just "CPU (NumPy)"
+3. No install hints shown when CuPy IS available
+
+---
+
+## Task 5: Build Environment Setup (Priority 1 — Manual by Олег)
+
+### Prerequisites for GPU build:
+
+```powershell
+# 1. Install CuPy in Windows Python (NOT WSL)
+pip install cupy-cuda12x
+
+# 2. Verify CuPy works
+python -c "import cupy; print(cupy.cuda.runtime.runtimeGetVersion())"
+# Should print: 12000 or similar
+
+# 3. Install PyInstaller if not present
+pip install pyinstaller
+
+# 4. Verify fastrlock (CuPy dependency)
+pip install fastrlock
+```
+
+### Build commands:
+
+```powershell
+# CPU-only build (existing)
+cd D:\root\rfcp\backend
+pyinstaller ..\installer\rfcp-server.spec --clean --noconfirm
+
+# GPU build (new)
+cd D:\root\rfcp\backend  
+pyinstaller ..\installer\rfcp-server-gpu.spec --clean --noconfirm
+```
+
+### Expected output sizes:
+```
+rfcp-server.exe (CPU):  ~80 MB
+rfcp-server.exe (GPU):  ~600-800 MB  (CuPy bundles CUDA runtime libs)
+```
+
+---
+
+## Task 6: Electron — Detect Build Variant (Priority 2 — 10 min)
+
+### File: `desktop/main.js` or `desktop/src/main.ts`
+
+Add version detection so UI knows which build it's running:
+
+```javascript
+// After backend starts, check GPU status
+async function checkBackendCapabilities() {
+  try {
+    const response = await fetch('http://127.0.0.1:8090/api/gpu/status');
+    const data = await response.json();
+    
+    // Send to renderer
+    mainWindow.webContents.send('gpu-status', data);
+    
+    if (data.available) {
+      console.log(`[RFCP] GPU: ${data.device} (${data.memory_mb} MB)`);
+    } else {
+      console.log('[RFCP] Running in CPU mode');
+    }
+  } catch (e) {
+    console.log('[RFCP] Backend not ready for GPU check');
+  }
+}
+```
+
+---
+
+## Task 7: About / Version Info (Priority 3 — 5 min)
+
+### Add build info to `/api/health` response:
+
+```python
+@app.get("/api/health")
+async def health():
+    gpu_info = get_gpu_info()
+    return {
+        "status": "ok",
+        "version": "3.6.0",
+        "build": "gpu" if gpu_info["available"] else "cpu",
+        "gpu": gpu_info,
+        "python": sys.version,
+        "platform": platform.platform(),
+    }
+```
+
+---
+
+## Build & Test Procedure
+
+### Step 1: Setup Build Environment
+```powershell
+# Windows PowerShell (NOT WSL)
+cd D:\root\rfcp
+
+# Verify Python environment  
+python --version       # Should be 3.11.x
+pip list | findstr cupy  # Should show cupy-cuda12x
+
+# If CuPy not installed:
+pip install cupy-cuda12x fastrlock
+```
+
+### Step 2: Build GPU Variant
+```powershell
+cd D:\root\rfcp\backend
+pyinstaller ..\installer\rfcp-server-gpu.spec --clean --noconfirm
+```
+
+### Step 3: Test Standalone
+```powershell
+# Run the built exe directly
+.\dist\rfcp-server.exe
+
+# In another terminal:
+curl http://localhost:8090/api/health
+curl http://localhost:8090/api/gpu/status
+curl http://localhost:8090/api/gpu/diagnostics
+```
+
+### Step 4: Verify GPU Detection
+Expected `/api/gpu/status` response:
+```json
+{
+  "available": true,
+  "backend": "CuPy (CUDA)",
+  "device": "NVIDIA GeForce RTX 4060 Laptop GPU",
+  "memory_mb": 8188
+}
+```
+
+### Step 5: Run Coverage Calculation
+- Place a site on map
+- Calculate coverage (10km, 200m resolution)
+- Check logs for: `[GPU] Using CUDA: RTX 4060 (8188 MB)`
+- Compare performance: should be 5-10x faster than CPU
+
+### Step 6: Full Electron Build
+```powershell
+# Copy GPU server to Electron resources
+copy backend\dist\rfcp-server.exe desktop\resources\
+
+# Build Electron installer
+cd installer
+.\build-win.sh  # or equivalent Windows script
+```
+
+---
+
+## Risk Assessment
+
+### Size Concern
+CuPy bundles CUDA runtime (~500MB). Total GPU installer ~700-800MB.
+**Mitigation:** This is acceptable for a professional RF planning tool.
+AutoCAD is 7GB. QGIS is 1.5GB. Atoll is 3GB+.
+
+### CUDA Version Compatibility
+CuPy-cuda12x requires CUDA 12.x compatible driver.
+RTX 4060 with Driver 581.42 → CUDA 13.0 → backward compatible ✅
+**Mitigation:** gpu_backend.py already falls back to NumPy gracefully.
+
+### PyInstaller + CuPy Issues
+Known issues:
+- CuPy uses many .so/.dll files that PyInstaller might miss
+- `collect_all('cupy')` should catch them, but test thoroughly
+- If missing DLLs → add them manually to `binaries` list
+
+**Mitigation:** Test the standalone exe on a clean machine (no Python installed).
+
+### Antivirus False Positives
+Larger exe = more AV suspicion. PyInstaller exes already trigger some AV.
+**Mitigation:** Code-sign the exe (future task), submit to AV vendors for whitelisting.
+
+---
+
+## Success Criteria
+
+- [ ] `rfcp-server-gpu.spec` created and builds successfully
+- [ ] Built exe detects RTX 4060 on startup
+- [ ] `/api/gpu/status` returns `"available": true`
+- [ ] Coverage calculation uses CuPy (check logs)
+- [ ] GPU badge shows "⚡ RTX 4060" (green) in header
+- [ ] Fallback to NumPy works if CUDA unavailable
+- [ ] CPU-only spec (`rfcp-server.spec`) still builds and works
+- [ ] Build time < 10 minutes
+- [ ] GPU exe size < 1 GB
+
+---
+
+## Commit Message
+
+```
+feat(build): add GPU-enabled PyInstaller build with CuPy + CUDA
+
+- New rfcp-server-gpu.spec with CuPy/CUDA collection
+- Build scripts: build-gpu.bat, build-gpu.sh
+- Graceful GPU detection in gpu_backend.py
+- Two-tier build: CPU (~80MB) and GPU (~700MB) variants
+- Auto-detection: RTX 4060 → CuPy acceleration
+- Fallback: no CUDA → NumPy (CPU mode)
+
+Iteration 3.6.0 — Production GPU Build
+```
+
+---
+
+## Files Summary
+
+### New Files:
+| File | Purpose |
+|------|---------|
+| `installer/rfcp-server-gpu.spec` | PyInstaller config with CuPy |
+| `installer/build-gpu.bat` | Windows GPU build script |
+| `installer/build-gpu.sh` | Linux/WSL GPU build script |
+
+### Modified Files:
+| File | Changes |
+|------|---------|
+| `backend/app/services/gpu_backend.py` | Verify graceful detection |
+| `backend/app/main.py` | Health endpoint with build info |
+| `desktop/main.js` or `main.ts` | GPU status check after backend start |
+| `frontend/src/components/ui/GPUIndicator.tsx` | Verify badge shows GPU |
+
+### No Changes Needed:
+| File | Reason |
+|------|--------|
+| `installer/rfcp-server.spec` | CPU build stays as-is |
+| `backend/app/services/coverage_service.py` | Already uses get_array_module() |
+| `installer/build-win.sh` | Existing CPU build unchanged |
+
+---
+
+## Timeline
+
+| Phase | Task | Time |
+|-------|------|------|
+| **P1** | Create rfcp-server-gpu.spec | 30 min |
+| **P1** | Build scripts | 15 min |
+| **P1** | Verify gpu_backend.py | 15 min |
+| **P2** | Frontend badge verification | 10 min |
+| **P2** | Electron GPU status | 10 min |
+| **P3** | Health endpoint update | 5 min |
+| **Test** | Build + test standalone | 20 min |
+| **Test** | Full Electron build | 15 min |
+| | **Total** | **~2 hours** |
+
+**Claude Code estimated time: 10-15 min** (spec + scripts + backend changes)
+**Manual testing by Олег: 30-45 min** (building + verifying)
--- a/UMTC-Wiki-MEGA-TASK.md
+++ b/UMTC-Wiki-MEGA-TASK.md
@@ -0,0 +1,352 @@
+# UMTC Wiki v2.0 — MEGA TASK: Integration & Polish
+
+Read UMTC-Wiki-v2.0-REFACTOR.md and UMTC-Wiki-v2.0-ROADMAP.md for full context.
+
+This is a comprehensive task covering all remaining fixes and integration work.
+Take your time, think hard, work through each section systematically.
+Report after completing each major section.
+
+---
+
+## SECTION A: Fix Critical Tauri 404 Bug
+
+The sidebar loads the full content tree correctly but clicking ANY article shows 404.
+
+### Debug steps:
+1. In `frontend/src/lib/api.ts` — find where getPage is called with a slug
+   Add `console.log('[WIKI] getPage called with slug:', slug)` 
+
+2. In `frontend/src/lib/utils/backend.ts` — in the tauriGetPage function
+   Add `console.log('[WIKI] Tauri invoke get_page with:', slug)`
+
+3. In `desktop/src-tauri/src/commands/content.rs` — in the get_page handler
+   Add `eprintln!("[WIKI] get_page received slug: {}", slug)`
+   Add `eprintln!("[WIKI] trying path: {:?}", resolved_path)`
+
+4. Check the Sidebar.svelte component — what href/slug does it generate when user clicks?
+   The web version uses `/api/pages/{slug}` — in desktop mode it should invoke with just the slug part.
+
+5. Common mismatches to check:
+   - Leading slash: sidebar sends `/lte/bbu` but Rust expects `lte/bbu`
+   - File extension: Rust looks for `lte/bbu.md` but file is `lte/bbu/index.md`
+   - URL encoding: Ukrainian characters in slugs
+   - The SvelteKit catch-all route `[...slug]` may pass the slug differently
+
+6. Fix the mismatch. Test navigation to at least 10 different pages including:
+   - Root sections (lte/, ran/, mikrotik/)
+   - Nested pages (lte/bbu, ran/srsenb-config)
+   - Glossary terms (glossary/prb)
+   - Deep nesting if any
+
+---
+
+## SECTION B: Fix Web Deployment
+
+The web version must keep working. Test and fix:
+
+1. Check that `backend/content.py` imports work:
+   - `from wiki_frontmatter import ArticleFrontmatter` 
+   - `from wikilinks import WikiLinksExtension`
+   - `from backlinks import BacklinksIndex`
+   - `from admonitions import AdmonitionsExtension`
+   
+   If any import fails, fix the module.
+
+2. Add the admonitions extension to the markdown pipeline in content.py
+   (wikilinks was already integrated, verify admonitions too)
+
+3. Make sure the backlinks API endpoint in main.py works:
+   - GET /api/pages/{slug:path}/backlinks
+   - Should return { "slug": "...", "backlinks": [...], "count": N }
+
+4. Add grade/status/category to the page API response:
+   - GET /api/pages/{slug} should now include grade, status, category fields
+
+5. Create a simple test script `scripts/test_web.py`:
+   ```python
+   # Test that backend starts and key endpoints work
+   import requests
+   BASE = "http://localhost:8000"
+   
+   # Test navigation
+   r = requests.get(f"{BASE}/api/navigation")
+   assert r.status_code == 200
+   nav = r.json()
+   print(f"Navigation: {len(nav)} sections")
+   
+   # Test page load
+   r = requests.get(f"{BASE}/api/pages/index")
+   assert r.status_code == 200
+   print(f"Home page: {r.json().get('title', 'OK')}")
+   
+   # Test search
+   r = requests.get(f"{BASE}/api/search?q=LTE")
+   assert r.status_code == 200
+   print(f"Search 'LTE': {len(r.json())} results")
+   
+   # Test backlinks
+   r = requests.get(f"{BASE}/api/pages/glossary/enb/backlinks")
+   print(f"Backlinks for eNB: {r.status_code}")
+   
+   print("\nAll tests passed!")
+   ```
+
+---
+
+## SECTION C: Frontend Wiki Components — Full Integration
+
+### C.1: Article Grade Badge on Pages
+
+In the wiki page view (`frontend/src/routes/[...slug]/+page.svelte` or equivalent):
+- Import ArticleGrade component
+- Display the grade badge next to the page title
+- The grade comes from the page API response (field: `grade`)
+- If no grade, don't show badge
+- Style: small badge inline with title, not a separate block
+
+### C.2: Breadcrumbs Component
+
+Create/update `frontend/src/lib/components/wiki/Breadcrumbs.svelte`:
+```svelte
+<!-- Example: Головна > LTE > BBU Setup -->
+<nav class="breadcrumbs">
+  <a href="/">Головна</a>
+  <span class="separator">/</span>
+  <a href="/lte">LTE</a>
+  <span class="separator">/</span>
+  <span class="current">BBU Setup</span>
+</nav>
+```
+- Generate from current page slug
+- Each segment is a link except the last
+- Use titles from navigation tree if available, otherwise humanize slug
+- Works in both web and desktop mode
+- Integrate into the page layout — show above article title
+
+### C.3: Admonition CSS
+
+Add styles for admonition boxes to the global CSS or a component:
+```css
+.admonition {
+  border-left: 4px solid;
+  border-radius: 4px;
+  padding: 12px 16px;
+  margin: 16px 0;
+}
+.admonition-note { border-color: #3b82f6; background: rgba(59,130,246,0.1); }
+.admonition-warning { border-color: #f59e0b; background: rgba(245,158,11,0.1); }
+.admonition-tip { border-color: #10b981; background: rgba(16,185,129,0.1); }
+.admonition-danger { border-color: #ef4444; background: rgba(239,68,68,0.1); }
+
+/* Dark mode */
+:global(.dark) .admonition-note { background: rgba(59,130,246,0.15); }
+:global(.dark) .admonition-warning { background: rgba(245,158,11,0.15); }
+:global(.dark) .admonition-tip { background: rgba(16,185,129,0.15); }
+:global(.dark) .admonition-danger { background: rgba(239,68,68,0.15); }
+
+.admonition-title {
+  font-weight: 600;
+  margin-bottom: 4px;
+}
+.admonition-icon {
+  margin-right: 8px;
+}
+```
+
+### C.4: Wiki-Link CSS
+
+Add styles for wiki-links:
+```css
+.wiki-link {
+  color: #3b82f6;
+  text-decoration: none;
+  border-bottom: 1px dotted #3b82f6;
+}
+.wiki-link:hover {
+  border-bottom-style: solid;
+}
+.red-link {
+  color: #ef4444;
+  border-bottom-color: #ef4444;
+}
+.red-link:hover::after {
+  content: " (сторінку не знайдено)";
+  font-size: 0.75em;
+  color: #9ca3af;
+}
+```
+
+### C.5: Backlinks Panel Integration
+
+In the page view, after the article content:
+- Show BacklinksPanel component
+- Pass current page slug
+- Works in both web (API) and desktop (Tauri IPC)
+- Only show if there are backlinks (don't show empty panel)
+
+### C.6: Table of Contents (sidebar)
+
+If the page has headings, generate a table of contents:
+- Extract h2/h3 from rendered HTML or use TOC data from backend
+- Show as a floating sidebar on wide screens (>1200px)
+- Collapsible on smaller screens
+- Highlight current section on scroll (intersection observer)
+- Works in both modes
+
+---
+
+## SECTION D: Search Integration for Desktop
+
+1. Test Tantivy search in Tauri:
+   - The search command should be wired to the Search component
+   - Type in search bar → results appear
+   - Cyrillic text should work (test: "мережа", "антена", "LTE")
+   
+2. If search doesn't work, debug:
+   - Is the search index built on startup? Check Rust logs
+   - Are content files found? Check content path resolution
+   - Is the query reaching the search command?
+
+3. Search results should show:
+   - Page title
+   - Brief excerpt (first 150 chars of content)
+   - Click navigates to page
+
+4. Keyboard shortcut: Ctrl+K should focus the search bar (already exists in web, verify in desktop)
+
+---
+
+## SECTION E: Content Quality Pass
+
+### E.1: Content Audit Script
+
+Create `scripts/analyze_content.py`:
+- Scan all .md files in content/
+- For each file report: has_frontmatter, word_count, has_code_blocks, grade, broken_wiki_links
+- Summary: total articles, by grade, articles needing work
+- Print actionable output
+
+### E.2: Add More Glossary Terms (20 more)
+
+Create glossary entries with proper frontmatter (grade: B, category: glossary):
+
+**Radio/RF terms:**
+- SGW (Serving Gateway)
+- PGW (PDN Gateway)  
+- HSS (Home Subscriber Server)
+- RSRP (Reference Signal Received Power)
+- RSRQ (Reference Signal Received Quality)
+- SINR (Signal to Interference plus Noise Ratio)
+- EARFCN (E-UTRA Absolute Radio Frequency Channel Number)
+- OFDM (Orthogonal Frequency Division Multiplexing)
+- MIMO (Multiple Input Multiple Output)
+- QoS (Quality of Service)
+
+**Infrastructure terms:**
+- WireGuard
+- MikroTik
+- Mesh Network
+- VLAN (Virtual LAN)
+- BGP (Border Gateway Protocol)
+- mTLS (Mutual TLS)
+- Caddy (Web Server)
+
+**Protocol terms:**
+- S1AP (S1 Application Protocol)
+- GTP (GPRS Tunnelling Protocol)
+- SCTP (Stream Control Transmission Protocol)
+
+Each glossary term should:
+- Have title in English with Ukrainian description
+- Use [[wiki-links]] to cross-reference other terms
+- Include: what it is, why it matters for UMTC, key parameters
+- Be 100-300 words
+
+### E.3: Upgrade 5 Key Articles to Grade B
+
+Pick the 5 most important articles and upgrade them:
+- Add proper frontmatter with all fields
+- Add :::note and :::warning admonitions where useful
+- Add [[wiki-links]] to glossary terms
+- Add "Див. також" (See also) section with related articles
+- Verify technical accuracy
+- Set grade: B
+
+Good candidates: 
+- Main LTE overview
+- srsENB configuration  
+- WireGuard setup
+- Open5GS overview
+- MikroTik basics
+
+---
+
+## SECTION F: Desktop App Polish
+
+### F.1: Window Title
+
+Show current page title in the window title bar:
+`UMTC Wiki — {Page Title}`
+
+### F.2: Keyboard Navigation
+
+- Arrow keys in sidebar to navigate
+- Enter to open selected item
+- Backspace to go back
+- Ctrl+K for search (verify)
+
+### F.3: Error Handling
+
+- If page not found, show a friendly Ukrainian message instead of generic 404
+- If content directory is missing, show setup instructions
+- If search index fails to build, log error but don't crash
+
+### F.4: About Dialog
+
+Add a simple about/info accessible from a gear icon or Help menu:
+- UMTC Wiki v2.0
+- Built with Tauri + SvelteKit + Rust
+- Content articles count
+- "Офлайн документація для УМТЗ"
+
+---
+
+## SECTION G: Production Builds
+
+### G.1: Web Docker Build Test
+
+Update docker-compose.yml if needed to include new backend modules.
+Make sure Dockerfile copies:
+- backend/wiki_frontmatter.py
+- backend/wikilinks.py  
+- backend/backlinks.py
+- backend/admonitions.py
+- All content/ files
+
+### G.2: Tauri Production Build
+
+Run `npx tauri build` and fix any remaining compilation errors.
+Report the output binary size and location.
+
+---
+
+## Order of Operations
+
+Do these in order — each section builds on the previous:
+
+1. **SECTION A** — Fix 404 bug (CRITICAL, everything depends on this)
+2. **SECTION B** — Verify web backend
+3. **SECTION C** — Frontend components
+4. **SECTION D** — Search
+5. **SECTION E** — Content
+6. **SECTION F** — Desktop polish
+7. **SECTION G** — Production builds
+
+Report after each section with:
+- What was done
+- What files were changed
+- Any issues found
+- Ready for next section?
+
+Think hard about edge cases. Don't break existing functionality.
+Good luck! 🚀
--- a/backend/app/api/routes/coverage.py
+++ b/backend/app/api/routes/coverage.py
@@ -14,6 +14,7 @@ from app.services.coverage_service import (
    select_propagation_model,
 )
 from app.services.parallel_coverage_service import CancellationToken
+from app.services.boundary_service import calculate_coverage_boundary

 router = APIRouter()

@@ -24,6 +25,12 @@ class CoverageRequest(BaseModel):
    settings: CoverageSettings = CoverageSettings()


+class BoundaryPoint(BaseModel):
+    """Single boundary coordinate"""
+    lat: float
+    lon: float
+
+
 class CoverageResponse(BaseModel):
    """Coverage calculation response"""
    points: List[CoveragePoint]
@@ -32,6 +39,7 @@ class CoverageResponse(BaseModel):
    stats: dict
    computation_time: float  # seconds
    models_used: List[str]  # which models were active
+    boundary: Optional[List[BoundaryPoint]] = None  # coverage boundary polygon


@router.post("/calculate")
@@ -131,13 +139,24 @@ async def calculate_coverage(request: CoverageRequest) -> CoverageResponse:
        "points_with_atmospheric_loss": sum(1 for p in points if p.atmospheric_loss > 0),
    }

+    # Calculate coverage boundary
+    boundary = None
+    if points:
+        boundary_coords = calculate_coverage_boundary(
+            [p.model_dump() for p in points],
+            threshold_dbm=request.settings.min_signal,
+        )
+        if boundary_coords:
+            boundary = [BoundaryPoint(**c) for c in boundary_coords]
+
    return CoverageResponse(
        points=points,
        count=len(points),
        settings=effective_settings,
        stats=stats,
        computation_time=round(computation_time, 2),
-        models_used=models_used
+        models_used=models_used,
+        boundary=boundary,
    )


--- a/backend/app/api/routes/health.py
+++ b/backend/app/api/routes/health.py
@@ -1,12 +1,29 @@
+import sys
+import platform
+
 from fastapi import APIRouter, Depends
 from app.api.deps import get_db
+from app.services.gpu_backend import gpu_manager

 router = APIRouter()


@router.get("/")
 async def health_check():
-    return {"status": "ok", "service": "rfcp-backend", "version": "1.1.0"}
+    gpu_info = gpu_manager.get_status()
+    return {
+        "status": "ok",
+        "service": "rfcp-backend",
+        "version": "3.6.0",
+        "build": "gpu" if gpu_info.get("gpu_available") else "cpu",
+        "gpu": {
+            "available": gpu_info.get("gpu_available", False),
+            "backend": gpu_info.get("active_backend", "cpu"),
+            "device": gpu_info.get("active_device", {}).get("name") if gpu_info.get("active_device") else "CPU",
+        },
+        "python": sys.version.split()[0],
+        "platform": platform.system(),
+    }


@router.get("/db")
--- a/backend/app/main.py
+++ b/backend/app/main.py
@@ -1,4 +1,6 @@
-from contextlib import asynccontextmanager
+from contextlib import asynccontextmanager
+import logging
+import platform

 from fastapi import FastAPI, WebSocket
 from fastapi.middleware.cors import CORSMiddleware
@@ -7,9 +9,54 @@ from app.core.database import connect_to_mongo, close_mongo_connection
 from app.api.routes import health, projects, terrain, coverage, regions, system, gpu
 from app.api.websocket import websocket_endpoint

+logger = logging.getLogger("rfcp.startup")
+
+
+def check_gpu_availability():
+    """Log GPU status on startup for debugging."""
+    is_wsl = "microsoft" in platform.release().lower()
+    env_note = " (WSL2)" if is_wsl else ""
+
+    # Check CuPy / CUDA
+    try:
+        import cupy as cp
+        device_count = cp.cuda.runtime.getDeviceCount()
+        if device_count > 0:
+            props = cp.cuda.runtime.getDeviceProperties(0)
+            name = props["name"]
+            if isinstance(name, bytes):
+                name = name.decode()
+            mem_mb = props["totalGlobalMem"] // (1024 * 1024)
+            logger.info(f"GPU detected{env_note}: {name} ({mem_mb} MB VRAM)")
+            logger.info(f"CuPy {cp.__version__}, CUDA devices: {device_count}")
+        else:
+            logger.warning(f"CuPy installed but no CUDA devices found{env_note}")
+    except Exception as e:
+        logger.warning(f"CuPy FAILED {env_note}: {e}")
+        if is_wsl:
+            logger.warning("Install: pip3 install cupy-cuda12x --break-system-packages")
+        else:
+            logger.warning("Install: pip install cupy-cuda12x")
+    except Exception as e:
+        logger.warning(f"CuPy error{env_note}: {e}")
+
+    # Check PyOpenCL
+    try:
+        import pyopencl as cl
+        platforms = cl.get_platforms()
+        for p in platforms:
+            for d in p.get_devices():
+                logger.info(f"OpenCL device: {d.name.strip()}")
+    except Exception as e:
+        logger.debug("PyOpenCL not installed (optional)")
+    except Exception:
+        pass
+

@asynccontextmanager
 async def lifespan(app: FastAPI):
+    # Log GPU status on startup
+    check_gpu_availability()
    await connect_to_mongo()
    yield
    await close_mongo_connection()
--- a/backend/app/services/boundary_service.py
+++ b/backend/app/services/boundary_service.py
@@ -0,0 +1,122 @@
+"""
+Coverage boundary calculation service.
+
+Computes concave hull (alpha shape) from coverage points to generate
+a realistic boundary that follows actual coverage contour.
+"""
+
+import logging
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+def calculate_coverage_boundary(
+    points: list[dict],
+    threshold_dbm: float = -100,
+    simplify_tolerance: float = 0.001,
+) -> list[dict]:
+    """
+    Calculate coverage boundary as concave hull of points above threshold.
+
+    Args:
+        points: List of coverage points with 'lat', 'lon', 'rsrp' keys
+        threshold_dbm: RSRP threshold - points below this are excluded
+        simplify_tolerance: Simplification tolerance in degrees (~100m per 0.001)
+
+    Returns:
+        List of {'lat': float, 'lon': float} coordinates forming boundary polygon.
+        Empty list if boundary cannot be computed.
+    """
+    try:
+        from shapely.geometry import MultiPoint
+        from shapely import concave_hull
+    except ImportError:
+        logger.warning("Shapely not installed - boundary calculation disabled")
+        return []
+
+    # Filter points above threshold
+    valid_coords = [
+        (p['lon'], p['lat'])  # Shapely uses (x, y) = (lon, lat)
+        for p in points
+        if p.get('rsrp', -999) >= threshold_dbm
+    ]
+
+    if len(valid_coords) < 3:
+        logger.debug(f"Not enough points for boundary: {len(valid_coords)}")
+        return []
+
+    try:
+        # Create MultiPoint geometry
+        mp = MultiPoint(valid_coords)
+
+        # Compute concave hull (alpha shape)
+        # ratio: 0 = convex hull, 1 = very tight fit
+        # 0.3-0.5 gives good balance between detail and smoothness
+        hull = concave_hull(mp, ratio=0.3)
+
+        if hull.is_empty:
+            logger.debug("Concave hull is empty")
+            return []
+
+        # Simplify to reduce points (0.001 deg ≈ 100m)
+        if simplify_tolerance > 0:
+            hull = hull.simplify(simplify_tolerance, preserve_topology=True)
+
+        # Extract coordinates based on geometry type
+        if hull.geom_type == 'Polygon':
+            coords = list(hull.exterior.coords)
+            return [{'lat': c[1], 'lon': c[0]} for c in coords]
+
+        elif hull.geom_type == 'MultiPolygon':
+            # Return largest polygon's exterior
+            largest = max(hull.geoms, key=lambda g: g.area)
+            coords = list(largest.exterior.coords)
+            return [{'lat': c[1], 'lon': c[0]} for c in coords]
+
+        elif hull.geom_type == 'GeometryCollection':
+            # Find polygons in collection
+            polygons = [g for g in hull.geoms if g.geom_type == 'Polygon']
+            if polygons:
+                largest = max(polygons, key=lambda g: g.area)
+                coords = list(largest.exterior.coords)
+                return [{'lat': c[1], 'lon': c[0]} for c in coords]
+
+        logger.debug(f"Unexpected hull geometry type: {hull.geom_type}")
+        return []
+
+    except Exception as e:
+        logger.warning(f"Boundary calculation error: {e}")
+        return []
+
+
+def calculate_multi_site_boundaries(
+    points: list[dict],
+    threshold_dbm: float = -100,
+) -> dict[str, list[dict]]:
+    """
+    Calculate separate boundaries for each site's coverage area.
+
+    Args:
+        points: Coverage points with 'lat', 'lon', 'rsrp', 'site_id' keys
+        threshold_dbm: RSRP threshold
+
+    Returns:
+        Dict mapping site_id to boundary coordinates list.
+    """
+    # Group points by site_id
+    by_site: dict[str, list[dict]] = {}
+    for p in points:
+        site_id = p.get('site_id', 'default')
+        if site_id not in by_site:
+            by_site[site_id] = []
+        by_site[site_id].append(p)
+
+    # Calculate boundary for each site
+    boundaries = {}
+    for site_id, site_points in by_site.items():
+        boundary = calculate_coverage_boundary(site_points, threshold_dbm)
+        if boundary:
+            boundaries[site_id] = boundary
+
+    return boundaries
--- a/backend/app/services/gpu_backend.py
+++ b/backend/app/services/gpu_backend.py
@@ -171,17 +171,34 @@ class GPUManager:
        """Full diagnostic info for troubleshooting GPU detection."""
        import sys
        import platform
+        import subprocess
+
+        is_wsl = "microsoft" in platform.release().lower()

        diag = {
            "python_version": sys.version,
+            "python_executable": sys.executable,
            "platform": platform.platform(),
+            "is_wsl": is_wsl,
            "numpy": {"version": np.__version__},
            "cuda": {},
            "opencl": {},
+            "nvidia_smi": None,
            "detected_devices": len(self._devices),
            "active_backend": self._active_backend.value,
        }

+        # Check nvidia-smi (works even without CuPy)
+        try:
+            result = subprocess.run(
+                ["nvidia-smi", "--query-gpu=name,memory.total,driver_version", "--format=csv,noheader"],
+                capture_output=True, text=True, timeout=5
+            )
+            if result.returncode == 0 and result.stdout.strip():
+                diag["nvidia_smi"] = result.stdout.strip()
+        except Exception:
+            diag["nvidia_smi"] = "not found or error"
+
        # Check CuPy/CUDA
        try:
            import cupy as cp
@@ -200,7 +217,10 @@ class GPUManager:
                }
        except ImportError:
            diag["cuda"]["error"] = "CuPy not installed"
-            diag["cuda"]["install_hint"] = "pip install cupy-cuda12x"
+            if is_wsl:
+                diag["cuda"]["install_hint"] = "pip3 install cupy-cuda12x --break-system-packages"
+            else:
+                diag["cuda"]["install_hint"] = "pip install cupy-cuda12x"
        except Exception as e:
            diag["cuda"]["error"] = str(e)

@@ -221,7 +241,10 @@ class GPUManager:
                diag["opencl"]["platforms"].append(platform_info)
        except ImportError:
            diag["opencl"]["error"] = "PyOpenCL not installed"
-            diag["opencl"]["install_hint"] = "pip install pyopencl"
+            if is_wsl:
+                diag["opencl"]["install_hint"] = "pip3 install pyopencl --break-system-packages"
+            else:
+                diag["opencl"]["install_hint"] = "pip install pyopencl"
        except Exception as e:
            diag["opencl"]["error"] = str(e)

--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -7,6 +7,7 @@ pymongo==4.6.1
 pydantic-settings==2.1.0
 numpy==1.26.4
 scipy==1.12.0
+shapely>=2.0.0
 requests==2.31.0
 httpx==0.27.0
 aiosqlite>=0.19.0
--- a/desktop/main.js
+++ b/desktop/main.js
@@ -52,9 +52,11 @@ const getLogPath = () => {
 const getBackendExePath = () => {
  const exeName = process.platform === 'win32' ? 'rfcp-server.exe' : 'rfcp-server';
  if (isDev) {
-    return path.join(__dirname, '..', 'backend', exeName);
+    // Dev: use the ONEDIR build output
+    return path.join(__dirname, '..', 'backend', 'dist', 'rfcp-server', exeName);
  }
-  return getResourcePath('backend', exeName);
+  // Production: ONEDIR structure - backend/rfcp-server/rfcp-server.exe
+  return getResourcePath('backend', 'rfcp-server', exeName);
 };

 /** Frontend index.html path (production only) */
--- a/docs/RFCP-Native-Backend-Research.md
+++ b/docs/RFCP-Native-Backend-Research.md
@@ -0,0 +1,233 @@
+# RFCP Native Backend Research
+
+## Executive Summary
+
+**Finding:** The production Electron app already supports native Windows operation without WSL2.
+
+The production build uses PyInstaller to bundle the Python backend as a standalone Windows executable (`rfcp-server.exe`). WSL2 is only used during development. No migration is needed for end users.
+
+---
+
+## Current Architecture
+
+### Development Mode
+```
+RFCP (Electron dev)
+  └── Spawns: python -m uvicorn app.main:app --host 127.0.0.1 --port 8090
+      └── Uses system Python (Windows or WSL2)
+          └── Requires venv with dependencies
+```
+
+### Production Mode (Already Implemented)
+```
+RFCP.exe (Electron packaged)
+  └── Spawns: rfcp-server.exe (bundled PyInstaller binary)
+      └── Self-contained Python + all dependencies
+      └── No WSL2 required
+      └── No system Python required
+```
+
+---
+
+## Evidence from Codebase
+
+### desktop/main.js (Lines 120-145)
+
+```javascript
+function startBackend() {
+  // Production: use bundled executable
+  if (isProduction) {
+    const serverPath = path.join(process.resourcesPath, 'rfcp-server.exe');
+    if (fs.existsSync(serverPath)) {
+      backendProcess = spawn(serverPath, [], { ... });
+      return;
+    }
+  }
+
+  // Development: use system Python
+  backendProcess = spawn('python', ['-m', 'uvicorn', 'app.main:app', ...]);
+}
+```
+
+### installer/rfcp-server.spec (PyInstaller Config)
+
+```python
+# Key configuration
+a = Analysis(
+    ['run_server.py'],
+    pathex=[backend_path],
+    binaries=[],
+    datas=[
+        ('data/terrain', 'data/terrain'),  # Terrain data bundled
+    ],
+    hiddenimports=[
+        'uvicorn.logging', 'uvicorn.loops', 'uvicorn.protocols',
+        'motor', 'pymongo', 'numpy', 'scipy', 'shapely',
+        # Full list of dependencies
+    ],
+)
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    name='rfcp-server',
+    console=True,  # Shows console for debugging
+    icon='rfcp.ico',
+)
+```
+
+---
+
+## GPU Acceleration in Production
+
+### Current Status
+The PyInstaller bundle **does not include CuPy** by default because:
+1. CuPy requires CUDA runtime (large, ~500MB)
+2. Not all users have NVIDIA GPUs
+3. Binary would be too large for distribution
+
+### Solution Options
+
+**Option A: Ship CPU-only (Current)**
+- Production build uses NumPy (CPU) for calculations
+- GPU acceleration available only in dev mode or manual install
+- Smallest download size (~100MB)
+
+**Option B: Separate GPU Installer**
+- Main installer: CPU-only (~100MB)
+- Optional GPU addon: Downloads CuPy + CUDA runtime (~600MB)
+- Implemented via install_rfcp.py dependency installer
+
+**Option C: CUDA Toolkit Detection**
+- Detect if CUDA is already installed on user's system
+- If yes, attempt to load CuPy dynamically
+- Graceful fallback to NumPy if not available
+
+### Recommendation
+Keep Option A (CPU-only production) with Option B available for power users:
+1. Default production build works everywhere
+2. Users with NVIDIA GPUs can run `install_rfcp.py` to enable GPU acceleration
+3. No WSL2 required for either path
+
+---
+
+## Terrain Data Handling
+
+### Current Implementation
+Terrain data (SRTM .hgt files) is bundled inside the PyInstaller executable:
+
+```python
+datas=[
+    ('data/terrain', 'data/terrain'),
+]
+```
+
+### Considerations
+- Bundled terrain data increases exe size significantly
+- Alternative: Download terrain on first use (like current region download system)
+- For initial release, bundling common regions is acceptable
+
+---
+
+## Database (MongoDB)
+
+### Production Architecture
+The Electron app embeds MongoDB or requires MongoDB to be installed separately.
+
+Options:
+1. **Embedded MongoDB** - Ships mongod.exe with the app
+2. **MongoDB Atlas** - Cloud database (requires internet)
+3. **SQLite** - Switch to file-based database (significant refactor)
+4. **In-memory + file persistence** - No MongoDB required (significant refactor)
+
+Current implementation uses Motor (async MongoDB driver). For true standalone operation, consider SQLite migration in future iteration.
+
+---
+
+## Build Process
+
+### Current Build Commands
+
+```bash
+# Build backend executable
+cd /mnt/d/root/rfcp/backend
+pyinstaller ../installer/rfcp-server.spec
+
+# Build Electron app with bundled backend
+cd /mnt/d/root/rfcp/installer
+./build-win.sh
+```
+
+### Output
+- `rfcp-server.exe` - Standalone backend (~80MB)
+- `RFCP-Setup-{version}.exe` - Full installer with Electron + backend (~150MB)
+
+---
+
+## Testing Native Build
+
+### Test Procedure
+1. Build `rfcp-server.exe` via PyInstaller
+2. Run directly: `./rfcp-server.exe`
+3. Verify API responds: `curl http://localhost:8090/api/health`
+4. Verify coverage calculation works
+5. Check GPU detection in logs
+
+### Known Issues
+1. **First launch slow**: PyInstaller extracts on first run (~5-10 seconds)
+2. **Antivirus false positives**: Some AV flags PyInstaller executables
+3. **Console window**: Shows black console (use `console=False` for windowless)
+
+---
+
+## Conclusions
+
+### No Migration Needed
+The production Electron app already works without WSL2. The current architecture is:
+- ✅ Native Windows executable
+- ✅ No Python installation required
+- ✅ No WSL2 required
+- ✅ Self-contained dependencies
+
+### Development vs Production
+| Aspect | Development | Production |
+|--------|-------------|------------|
+| Python | System Python / venv | Bundled via PyInstaller |
+| WSL2 | Optional (for testing) | Not required |
+| GPU | CuPy if installed | CPU-only (NumPy) |
+| MongoDB | Local instance | Embedded or Atlas |
+| Terrain | Local data/ folder | Bundled in exe |
+
+### Remaining Work
+1. **GPU for production**: Implement Optional GPU addon installer
+2. **Smaller package**: On-demand terrain download instead of bundling
+3. **Database portability**: Consider SQLite migration for offline-first
+4. **Installer polish**: Signed executables, auto-update support
+
+---
+
+## Appendix: Full PyInstaller Hidden Imports
+
+From `installer/rfcp-server.spec`:
+```python
+hiddenimports=[
+    'uvicorn.logging',
+    'uvicorn.loops',
+    'uvicorn.loops.auto',
+    'uvicorn.protocols',
+    'uvicorn.protocols.http',
+    'uvicorn.protocols.http.auto',
+    'uvicorn.protocols.websockets',
+    'uvicorn.protocols.websockets.auto',
+    'uvicorn.lifespan',
+    'uvicorn.lifespan.on',
+    'motor',
+    'pymongo',
+    'numpy',
+    'scipy',
+    'shapely',
+    'shapely.geometry',
+    'shapely.ops',
+    # ... additional imports
+]
+```
--- a/frontend/src/App.tsx
+++ b/frontend/src/App.tsx
@@ -444,11 +444,14 @@ export default function App() {
          );
        } else {
          const timeStr = result.calculationTime.toFixed(1);
+          const firstSite = sites.find((s) => s.visible);
+          const freqStr = firstSite ? ` \u2022 ${firstSite.frequency} MHz` : '';
+          const presetStr = settings.preset ? ` \u2022 ${settings.preset}` : '';
          const modelsStr = result.modelsUsed?.length
-            ? ` • ${result.modelsUsed.length} models`
+            ? ` \u2022 ${result.modelsUsed.length} models`
            : '';
          addToast(
-            `Calculated ${result.totalPoints.toLocaleString()} points in ${timeStr}s${modelsStr}`,
+            `${result.totalPoints.toLocaleString()} pts \u2022 ${timeStr}s${presetStr}${freqStr}${modelsStr}`,
            'success'
          );
        }
@@ -481,7 +484,7 @@ export default function App() {
  return (
    <div className="h-screen w-screen flex flex-col bg-gray-100 dark:bg-dark-bg">
      {/* Header */}
-      <header className="bg-slate-800 dark:bg-slate-900 text-white px-4 py-2 flex items-center justify-between flex-shrink-0 z-10">
+      <header className="bg-slate-800 dark:bg-slate-900 text-white px-4 py-2 flex items-center justify-between flex-shrink-0 z-[1010]">
        <div className="flex items-center gap-2">
          <span className="text-base font-bold">RFCP</span>
          <span className="text-xs text-slate-400 hidden sm:inline">
@@ -684,6 +687,7 @@ export default function App() {
                    points={coverageResult.points.filter(p => p.rsrp >= settings.rsrpThreshold)}
                    visible={showBoundary}
                    resolution={settings.resolution}
+                    boundary={coverageResult.boundary}
                  />
                )}
              </>
--- a/frontend/src/components/map/CoverageBoundary.tsx
+++ b/frontend/src/components/map/CoverageBoundary.tsx
@@ -1,8 +1,8 @@
 /**
 * Renders a dashed polyline around the coverage zone boundary.
 *
- * Uses @turf/concave to compute a concave hull (alpha shape) per site,
- * which correctly follows sector/wedge shapes — not just convex circles.
+ * Prefers server-computed boundary if available (shapely concave_hull).
+ * Falls back to client-side @turf/concave computation.
 *
 * Performance: ~20-50ms for 10k points (runs once per coverage change).
 */
@@ -12,7 +12,7 @@ import { useMap } from 'react-leaflet';
 import L from 'leaflet';
 import concave from '@turf/concave';
 import { featureCollection, point } from '@turf/helpers';
-import type { CoveragePoint } from '@/types/index.ts';
+import type { CoveragePoint, BoundaryPoint } from '@/types/index.ts';
 import { logger } from '@/utils/logger.ts';

 interface CoverageBoundaryProps {
@@ -21,6 +21,7 @@ interface CoverageBoundaryProps {
  resolution: number; // meters — controls concave hull detail
  color?: string;
  weight?: number;
+  boundary?: BoundaryPoint[]; // server-provided boundary (preferred)
 }

 export default function CoverageBoundary({
@@ -29,13 +30,25 @@ export default function CoverageBoundary({
  resolution,
  color = '#ffffff', // white — visible against red-to-blue gradient
  weight = 2,
+  boundary,
 }: CoverageBoundaryProps) {
  const map = useMap();
  const layerRef = useRef<L.LayerGroup | null>(null);

-  // Compute boundary paths grouped by site
+  // Compute boundary paths - prefer server boundary, fallback to client-side
  const boundaryPaths = useMemo(() => {
-    if (!visible || points.length === 0) return [];
+    if (!visible) return [];
+
+    // Use server-provided boundary if available
+    if (boundary && boundary.length >= 3) {
+      const serverPath: L.LatLngExpression[] = boundary.map(
+        (p) => [p.lat, p.lon] as L.LatLngExpression
+      );
+      return [serverPath];
+    }
+
+    // Fallback to client-side computation
+    if (points.length === 0) return [];

    // Group points by siteId (fallback to 'all' when siteId not available from API)
    const bySite = new Map<string, CoveragePoint[]>();
@@ -61,7 +74,7 @@ export default function CoverageBoundary({
    }

    return paths;
-  }, [points, visible, resolution]);
+  }, [points, visible, resolution, boundary]);

  // Render / cleanup polylines
  useEffect(() => {
--- a/frontend/src/components/map/MeasurementTool.tsx
+++ b/frontend/src/components/map/MeasurementTool.tsx
@@ -121,6 +121,7 @@ export default function MeasurementTool({ enabled, onComplete, onProfileRequest
            <button
              onClick={(e) => {
                e.stopPropagation();
+                e.preventDefault();
                onProfileRequest(points[0], points[points.length - 1]);
              }}
              style={{
--- a/frontend/src/components/ui/GPUIndicator.tsx
+++ b/frontend/src/components/ui/GPUIndicator.tsx
@@ -33,6 +33,13 @@ export default function GPUIndicator() {
    return () => document.removeEventListener('mousedown', handler);
  }, [open]);

+  // Auto-fetch diagnostics when dropdown opens and only CPU available
+  useEffect(() => {
+    if (open && status?.active_backend === 'cpu' && !diagnostics) {
+      api.getGPUDiagnostics().then(setDiagnostics).catch(() => {});
+    }
+  }, [open, status?.active_backend, diagnostics]);
+
  if (!status) return null;

  const isGPU = status.active_backend !== 'cpu';
@@ -119,15 +126,30 @@ export default function GPUIndicator() {
              <div className="text-[10px] text-yellow-600 dark:text-yellow-400 mb-2">
                No GPU detected. For faster calculations:
              </div>
-              <div className="text-[10px] text-gray-500 dark:text-dark-muted space-y-0.5">
-                <div>NVIDIA: <code className="bg-gray-100 dark:bg-dark-border px-1 rounded">pip install cupy-cuda12x</code></div>
-                <div>Intel/AMD: <code className="bg-gray-100 dark:bg-dark-border px-1 rounded">pip install pyopencl</code></div>
-              </div>
+              {diagnostics?.is_wsl ? (
+                <div className="text-[10px] text-gray-500 dark:text-dark-muted space-y-1">
+                  <div className="text-[9px] text-gray-400 dark:text-dark-muted mb-1">WSL2 detected - use pip3:</div>
+                  <div className="bg-gray-100 dark:bg-dark-border px-2 py-1 rounded font-mono text-[9px] break-all">
+                    pip3 install cupy-cuda12x --break-system-packages
+                  </div>
+                  <div className="text-[9px] text-gray-400 dark:text-dark-muted mt-1">Then restart RFCP</div>
+                </div>
+              ) : (
+                <div className="text-[10px] text-gray-500 dark:text-dark-muted space-y-0.5">
+                  <div>NVIDIA: <code className="bg-gray-100 dark:bg-dark-border px-1 rounded">pip install cupy-cuda12x</code></div>
+                  <div>Intel/AMD: <code className="bg-gray-100 dark:bg-dark-border px-1 rounded">pip install pyopencl</code></div>
+                </div>
+              )}
+              {typeof diagnostics?.nvidia_smi === 'string' && diagnostics.nvidia_smi !== 'not found or error' && (
+                <div className="mt-2 text-[9px] text-green-600 dark:text-green-400">
+                  GPU hardware found: {diagnostics.nvidia_smi.split(',')[0]}
+                </div>
+              )}
              <button
                onClick={handleRunDiagnostics}
                className="mt-2 w-full text-[10px] text-blue-600 dark:text-blue-400 hover:underline text-left"
              >
-                Run Diagnostics
+                {diagnostics ? 'Refresh Diagnostics' : 'Run Diagnostics'}
              </button>
            </div>
          )}
--- a/frontend/src/services/api.ts
+++ b/frontend/src/services/api.ts
@@ -75,6 +75,11 @@ export interface ApiCoverageStats {
  points_with_atmospheric_loss: number;
 }

+export interface ApiBoundaryPoint {
+  lat: number;
+  lon: number;
+}
+
 export interface CoverageResponse {
  points: ApiCoveragePoint[];
  count: number;
@@ -82,6 +87,7 @@ export interface CoverageResponse {
  stats: ApiCoverageStats;
  computation_time: number;
  models_used: string[];
+  boundary?: ApiBoundaryPoint[];
 }

 export interface Preset {
--- a/frontend/src/store/coverage.ts
+++ b/frontend/src/store/coverage.ts
@@ -98,6 +98,7 @@ function responseToResult(response: CoverageResponse, settings: CoverageSettings
    settings: settings,
    stats: response.stats as CoverageApiStats,
    modelsUsed: response.models_used,
+    boundary: response.boundary,
  };
 }

@@ -251,11 +252,14 @@ export const useCoverageStore = create<CoverageState>((set, get) => ({
              addToast('No coverage points. Try increasing radius.', 'warning');
            } else {
              const timeStr = result.calculationTime.toFixed(1);
+              const firstSite = useSitesStore.getState().sites.find((s) => s.visible);
+              const freqStr = firstSite ? ` \u2022 ${firstSite.frequency} MHz` : '';
+              const presetStr = settings.preset ? ` \u2022 ${settings.preset}` : '';
              const modelsStr = result.modelsUsed?.length
                ? ` \u2022 ${result.modelsUsed.length} models`
                : '';
              addToast(
-                `Calculated ${result.totalPoints.toLocaleString()} points in ${timeStr}s${modelsStr}`,
+                `${result.totalPoints.toLocaleString()} pts \u2022 ${timeStr}s${presetStr}${freqStr}${modelsStr}`,
                'success'
              );
            }
--- a/frontend/src/types/coverage.ts
+++ b/frontend/src/types/coverage.ts
@@ -15,6 +15,11 @@ export interface CoveragePoint {
  atmospheric_loss?: number;   // dB atmospheric absorption
 }

+export interface BoundaryPoint {
+  lat: number;
+  lon: number;
+}
+
 export interface CoverageResult {
  points: CoveragePoint[];
  calculationTime: number;     // seconds (was ms for browser calc)
@@ -23,6 +28,7 @@ export interface CoverageResult {
  // API-provided fields
  stats?: CoverageApiStats;
  modelsUsed?: string[];
+  boundary?: BoundaryPoint[];  // server-computed coverage boundary
 }

 export interface CoverageApiStats {
--- a/frontend/src/types/index.ts
+++ b/frontend/src/types/index.ts
@@ -5,5 +5,6 @@ export type {
  CoverageSettings,
  CoverageApiStats,
  GridPoint,
+  BoundaryPoint,
 } from './coverage.ts';
 export type { FrequencyBand } from './frequency.ts';
--- a/installer/build-gpu.bat
+++ b/installer/build-gpu.bat
@@ -0,0 +1,70 @@
+@echo off
+echo ========================================
+echo  RFCP GPU Build — ONEDIR mode
+echo  CuPy-cuda13x + CUDA Toolkit 13.x
+echo ========================================
+echo.
+
+REM ── Check CuPy ──
+echo [1/5] Checking CuPy installation...
+python -c "import cupy; print(f'  CuPy {cupy.__version__}')" 2>nul
+if errorlevel 1 (
+    echo ERROR: CuPy not installed.
+    echo   Run: pip install cupy-cuda13x
+    exit /b 1
+)
+
+REM ── Check CUDA compute ──
+echo [2/5] Testing GPU compute...
+python -c "import cupy; a = cupy.array([1,2,3]); assert a.sum() == 6; print('  GPU compute: OK')" 2>nul
+if errorlevel 1 (
+    echo ERROR: CuPy installed but GPU compute failed.
+    echo   Check: CUDA Toolkit installed? nvidia-smi works?
+    exit /b 1
+)
+
+REM ── Check CUDA_PATH ──
+echo [3/5] Checking CUDA Toolkit...
+if defined CUDA_PATH (
+    echo   CUDA_PATH: %CUDA_PATH%
+) else (
+    echo   WARNING: CUDA_PATH not set
+)
+
+REM ── Check nvidia pip DLLs ──
+echo [4/5] Checking nvidia pip packages...
+python -c "import nvidia; import os; base=os.path.dirname(nvidia.__file__); dlls=[f for d in os.listdir(base) if os.path.isdir(os.path.join(base,d,'bin')) for f in os.listdir(os.path.join(base,d,'bin')) if f.endswith('.dll')]; print(f'  nvidia pip DLLs: {len(dlls)}')" 2>nul
+if errorlevel 1 (
+    echo   No nvidia pip packages (will use CUDA Toolkit)
+)
+
+REM ── Build ──
+echo.
+echo [5/5] Building rfcp-server (ONEDIR mode)...
+echo   This may take 3-5 minutes...
+echo.
+
+cd /d "%~dp0\..\backend"
+pyinstaller "..\installer\rfcp-server-gpu.spec" --clean --noconfirm
+
+echo.
+echo ========================================
+if exist "dist\rfcp-server\rfcp-server.exe" (
+    echo  BUILD COMPLETE! (ONEDIR mode)
+    echo.
+    echo  Output: backend\dist\rfcp-server\
+    dir /b dist\rfcp-server\*.exe dist\rfcp-server\*.dll 2>nul | find /c /v "" > nul
+    echo.
+    echo  Test commands:
+    echo    cd dist\rfcp-server
+    echo    rfcp-server.exe
+    echo    curl http://localhost:8090/api/health
+    echo    curl http://localhost:8090/api/gpu/status
+    echo ========================================
+) else (
+    echo  BUILD FAILED — check errors above
+    echo ========================================
+    exit /b 1
+)
+
+pause
--- a/installer/build-gpu.sh
+++ b/installer/build-gpu.sh
@@ -0,0 +1,84 @@
+#!/bin/bash
+set -e
+
+echo "========================================"
+echo " RFCP GPU Build — ONEDIR mode"
+echo " CuPy-cuda13x + CUDA Toolkit 13.x"
+echo "========================================"
+echo ""
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+BACKEND_DIR="$SCRIPT_DIR/../backend"
+
+# Check backend exists
+if [ ! -f "$BACKEND_DIR/run_server.py" ]; then
+    echo "ERROR: Backend not found at $BACKEND_DIR"
+    exit 1
+fi
+
+# Check Python
+echo "[1/5] Checking Python..."
+python3 --version || { echo "ERROR: Python3 not found"; exit 1; }
+
+# Check CuPy
+echo ""
+echo "[2/5] Checking CuPy installation..."
+if ! python3 -c "import cupy; print(f'  CuPy {cupy.__version__}')" 2>/dev/null; then
+    echo "ERROR: CuPy not installed"
+    echo ""
+    echo "Install CuPy:"
+    echo "  pip3 install cupy-cuda13x"
+    echo "  # or for WSL2:"
+    echo "  pip3 install cupy-cuda13x --break-system-packages"
+    exit 1
+fi
+
+# Check GPU compute
+echo ""
+echo "[3/5] Testing GPU compute..."
+if python3 -c "import cupy; a = cupy.array([1,2,3]); assert a.sum() == 6; print('  GPU compute: OK')" 2>/dev/null; then
+    :
+else
+    echo "WARNING: GPU compute test failed (may still work)"
+fi
+
+# Check CUDA
+echo ""
+echo "[4/5] Checking CUDA..."
+if [ -n "$CUDA_PATH" ]; then
+    echo "  CUDA_PATH: $CUDA_PATH"
+else
+    echo "  CUDA_PATH not set (relying on nvidia pip packages)"
+fi
+
+# Check nvidia pip packages
+echo ""
+echo "[5/5] Checking nvidia pip packages..."
+python3 -c "import nvidia; print('  nvidia packages found')" 2>/dev/null || echo "  No nvidia pip packages"
+
+# Build
+echo ""
+echo "Building rfcp-server (ONEDIR mode)..."
+echo ""
+
+cd "$BACKEND_DIR"
+pyinstaller "$SCRIPT_DIR/rfcp-server-gpu.spec" --clean --noconfirm
+
+echo ""
+echo "========================================"
+if [ -f "dist/rfcp-server/rfcp-server" ] || [ -f "dist/rfcp-server/rfcp-server.exe" ]; then
+    echo " BUILD COMPLETE! (ONEDIR mode)"
+    echo ""
+    echo " Output: backend/dist/rfcp-server/"
+    ls -lh dist/rfcp-server/ | head -20
+    echo ""
+    echo " Test:"
+    echo "   cd dist/rfcp-server"
+    echo "   ./rfcp-server"
+    echo "   curl http://localhost:8090/api/health"
+    echo "========================================"
+else
+    echo " BUILD FAILED — check errors above"
+    echo "========================================"
+    exit 1
+fi
--- a/installer/build-win.sh
+++ b/installer/build-win.sh
@@ -3,6 +3,7 @@ set -e

 echo "========================================="
 echo "  RFCP Desktop Build (Windows)"
+echo "  GPU-enabled ONEDIR build"
 echo "========================================="

 cd "$(dirname "$0")/.."
@@ -14,15 +15,30 @@ npm ci
 npm run build
 cd ..

-# 2. Build backend with PyInstaller
-echo "[2/4] Building backend..."
+# 2. Build backend with PyInstaller (GPU ONEDIR mode)
+echo "[2/4] Building backend (GPU)..."
 cd backend
+
+# Check CuPy is available
+if ! python -c "import cupy" 2>/dev/null; then
+    echo "WARNING: CuPy not installed - GPU acceleration will not be available"
+    echo "  Install with: pip install cupy-cuda13x"
+fi
+
 python -m pip install -r requirements.txt
 python -m pip install pyinstaller
-cd ../installer
-python -m PyInstaller rfcp-server.spec --clean --noconfirm
+
+# Build using GPU spec (ONEDIR output)
+python -m PyInstaller ../installer/rfcp-server-gpu.spec --clean --noconfirm
+
+# Copy ONEDIR folder to desktop staging area
+# Result: desktop/backend-dist/win/rfcp-server/rfcp-server.exe + _internal/
 mkdir -p ../desktop/backend-dist/win
-cp dist/rfcp-server.exe ../desktop/backend-dist/win/
+rm -rf ../desktop/backend-dist/win/rfcp-server  # Clean old build
+cp -r dist/rfcp-server ../desktop/backend-dist/win/rfcp-server
+
+echo "  Backend copied to: desktop/backend-dist/win/rfcp-server/"
+ls -la ../desktop/backend-dist/win/rfcp-server/*.exe 2>/dev/null || true
 cd ..

 # 3. Build Electron app
--- a/installer/rfcp-server-gpu.spec
+++ b/installer/rfcp-server-gpu.spec
@@ -0,0 +1,305 @@
+# rfcp-server-gpu.spec — GPU-enabled build (CuPy + CUDA 13.x)
+# RFCP Iteration 3.6.0
+#
+# Mode: ONEDIR (directory output, not single exe)
+# This is better for CUDA — DLLs load directly without temp extraction
+#
+# Requirements:
+#   pip install cupy-cuda13x fastrlock pyinstaller
+#   CUDA Toolkit 13.x installed (winget install Nvidia.CUDA)
+#
+# Build:
+#   cd backend && pyinstaller ../installer/rfcp-server-gpu.spec --clean --noconfirm
+#
+# Output:
+#   backend/dist/rfcp-server/rfcp-server.exe  (+ DLLs in same folder)
+
+import os
+import sys
+import glob
+from PyInstaller.utils.hooks import collect_all, collect_dynamic_libs
+
+backend_path = os.path.abspath(os.path.join(os.path.dirname(SPEC), '..', 'backend'))
+print(f"[GPU SPEC] Backend path: {backend_path}")
+
+# ═══════════════════════════════════════════
+# Collect CuPy packages
+# ═══════════════════════════════════════════
+cupy_datas = []
+cupy_binaries = []
+cupy_hiddenimports = []
+cupyb_datas = []
+cupyb_binaries = []
+cupyb_hiddenimports = []
+
+try:
+    cupy_datas, cupy_binaries, cupy_hiddenimports = collect_all('cupy')
+    cupyb_datas, cupyb_binaries, cupyb_hiddenimports = collect_all('cupy_backends')
+    print(f"[GPU SPEC] CuPy: {len(cupy_binaries)} binaries, {len(cupy_datas)} data files")
+except Exception as e:
+    print(f"[GPU SPEC] WARNING: CuPy collection failed: {e}")
+
+# NOTE: nvidia pip packages REMOVED - they have cuda12 DLLs that conflict with cupy-cuda13x
+# We use CUDA Toolkit 13.x DLLs only
+
+# ═══════════════════════════════════════════
+# Collect CUDA Toolkit DLLs (system install)
+# ═══════════════════════════════════════════
+# Installed via: winget install Nvidia.CUDA
+cuda_toolkit_binaries = []
+cuda_path = os.environ.get('CUDA_PATH', '')
+
+if cuda_path:
+    # Scan BOTH bin\ and bin\x64\ directories
+    cuda_bin_dirs = [
+        os.path.join(cuda_path, 'bin'),
+        os.path.join(cuda_path, 'bin', 'x64'),
+    ]
+
+    # Only essential CUDA runtime DLLs (exclude NPP, nvjpeg, nvblas, nvfatbin)
+    cuda_dll_patterns = [
+        'cublas64_*.dll',
+        'cublasLt64_*.dll',
+        'cudart64_*.dll',
+        'cufft64_*.dll',
+        'cufftw64_*.dll',
+        'curand64_*.dll',
+        'cusolver64_*.dll',
+        'cusolverMg64_*.dll',
+        'cusparse64_*.dll',
+        'nvrtc64_*.dll',
+        'nvrtc-builtins64_*.dll',
+        'nvJitLink_*.dll',
+        'nvjitlink_*.dll',
+    ]
+
+    collected_dlls = set()  # Avoid duplicates
+    for cuda_bin in cuda_bin_dirs:
+        if os.path.isdir(cuda_bin):
+            for pattern in cuda_dll_patterns:
+                for dll in glob.glob(os.path.join(cuda_bin, pattern)):
+                    dll_name = os.path.basename(dll)
+                    if dll_name not in collected_dlls:
+                        cuda_toolkit_binaries.append((dll, '.'))
+                        collected_dlls.add(dll_name)
+            print(f"[GPU SPEC] Scanned: {cuda_bin}")
+
+    print(f"[GPU SPEC] CUDA Toolkit ({cuda_path}): {len(cuda_toolkit_binaries)} DLLs")
+    for dll, _ in cuda_toolkit_binaries:
+        print(f"[GPU SPEC]   {os.path.basename(dll)}")
+else:
+    print("[GPU SPEC] ERROR: CUDA_PATH not set!")
+    print("[GPU SPEC]   Install: winget install Nvidia.CUDA")
+
+# All GPU binaries (CUDA Toolkit only, no nvidia pip packages)
+all_gpu_binaries = cuda_toolkit_binaries
+
+if len(all_gpu_binaries) == 0:
+    print("[GPU SPEC] ⚠ NO CUDA DLLs FOUND!")
+    print("[GPU SPEC]   Install CUDA Toolkit: winget install Nvidia.CUDA")
+else:
+    print(f"[GPU SPEC] ✅ Total GPU DLLs: {len(all_gpu_binaries)}")
+
+# ═══════════════════════════════════════════
+# Collect fastrlock (CuPy dependency)
+# ═══════════════════════════════════════════
+fl_datas = []
+fl_binaries = []
+fl_hiddenimports = []
+try:
+    fl_datas, fl_binaries, fl_hiddenimports = collect_all('fastrlock')
+    print(f"[GPU SPEC] fastrlock: {len(fl_binaries)} binaries")
+except Exception:
+    print("[GPU SPEC] fastrlock not found (optional)")
+
+# ═══════════════════════════════════════════
+# PyInstaller Analysis
+# ═══════════════════════════════════════════
+
+a = Analysis(
+    [os.path.join(backend_path, 'run_server.py')],
+    pathex=[backend_path],
+    binaries=(
+        cupy_binaries + cupyb_binaries +
+        fl_binaries + all_gpu_binaries
+    ),
+    datas=[
+        # Include app/ source code
+        (os.path.join(backend_path, 'app'), 'app'),
+    ] + cupy_datas + cupyb_datas + fl_datas,
+    hiddenimports=[
+        # ── Uvicorn internals ──
+        'uvicorn.logging',
+        'uvicorn.loops',
+        'uvicorn.loops.auto',
+        'uvicorn.loops.asyncio',
+        'uvicorn.protocols',
+        'uvicorn.protocols.http',
+        'uvicorn.protocols.http.auto',
+        'uvicorn.protocols.http.h11_impl',
+        'uvicorn.protocols.http.httptools_impl',
+        'uvicorn.protocols.websockets',
+        'uvicorn.protocols.websockets.auto',
+        'uvicorn.protocols.websockets.wsproto_impl',
+        'uvicorn.lifespan',
+        'uvicorn.lifespan.on',
+        'uvicorn.lifespan.off',
+        # ── FastAPI / Starlette ──
+        'fastapi',
+        'fastapi.middleware',
+        'fastapi.middleware.cors',
+        'fastapi.routing',
+        'fastapi.responses',
+        'fastapi.exceptions',
+        'starlette',
+        'starlette.routing',
+        'starlette.middleware',
+        'starlette.middleware.cors',
+        'starlette.responses',
+        'starlette.requests',
+        'starlette.concurrency',
+        'starlette.formparsers',
+        'starlette.staticfiles',
+        # ── Pydantic ──
+        'pydantic',
+        'pydantic.fields',
+        'pydantic_settings',
+        'pydantic_core',
+        # ── HTTP / networking ──
+        'httpx',
+        'httpcore',
+        'h11',
+        'httptools',
+        'anyio',
+        'anyio._backends',
+        'anyio._backends._asyncio',
+        'sniffio',
+        # ── MongoDB (motor/pymongo) ──
+        'motor',
+        'motor.motor_asyncio',
+        'pymongo',
+        'pymongo.errors',
+        'pymongo.collection',
+        'pymongo.database',
+        'pymongo.mongo_client',
+        # ── Async I/O ──
+        'aiofiles',
+        'aiofiles.os',
+        'aiofiles.ospath',
+        # ── Scientific ──
+        'numpy',
+        'numpy.core',
+        'scipy',
+        'scipy.special',
+        'scipy.interpolate',
+        'shapely',
+        'shapely.geometry',
+        'shapely.ops',
+        # ── Multipart ──
+        'multipart',
+        'python_multipart',
+        # ── Encoding ──
+        'email.mime',
+        'email.mime.multipart',
+        # ── Multiprocessing ──
+        'multiprocessing',
+        'multiprocessing.pool',
+        'multiprocessing.queues',
+        'concurrent.futures',
+        # ── CuPy + CUDA ──
+        'cupy',
+        'cupy.cuda',
+        'cupy.cuda.runtime',
+        'cupy.cuda.driver',
+        'cupy.cuda.memory',
+        'cupy.cuda.stream',
+        'cupy.cuda.device',
+        'cupy._core',
+        'cupy._core.core',
+        'cupy._core._routines_math',
+        'cupy._core._routines_logic',
+        'cupy._core._routines_manipulation',
+        'cupy._core._routines_sorting',
+        'cupy._core._routines_statistics',
+        'cupy._core._cub_reduction',
+        'cupy.fft',
+        'cupy.linalg',
+        'cupy.random',
+        'cupy_backends',
+        'cupy_backends.cuda',
+        'cupy_backends.cuda.api',
+        'cupy_backends.cuda.libs',
+        'fastrlock',
+        'fastrlock.rlock',
+    ] + cupy_hiddenimports + cupyb_hiddenimports + fl_hiddenimports,
+    hookspath=[],
+    hooksconfig={},
+    runtime_hooks=[os.path.join(os.path.dirname(SPEC), 'rthook_cuda_dlls.py')],
+    # ── Exclude bloat ──
+    excludes=[
+        # GUI
+        'tkinter',
+        'matplotlib',
+        'PIL',
+        'IPython',
+        # Data science bloat
+        'pandas',
+        'tensorflow',
+        'torch',
+        'keras',
+        # Testing
+        'pytest',
+        
+        # Jupyter
+        'jupyter',
+        'notebook',
+        'ipykernel',
+        # gRPC / telemetry (often pulled in by dependencies)
+        'grpc',
+        'grpcio',
+        'google.protobuf',
+        'opentelemetry',
+        'opentelemetry.sdk',
+        'opentelemetry.instrumentation',
+        # Ray (too heavy, we use multiprocessing)
+        'ray',
+        # Other
+        'cv2',
+        'sklearn',
+        'sympy',
+    ],
+    noarchive=False,
+)
+
+pyz = PYZ(a.pure)
+
+# ═══════════════════════════════════════════
+# ONEDIR mode: EXE + COLLECT
+# ═══════════════════════════════════════════
+# Creates: dist/rfcp-server/rfcp-server.exe + all DLLs in same folder
+# Better for CUDA — no temp extraction needed
+
+exe = EXE(
+    pyz,
+    a.scripts,
+    [],  # No binaries/datas in EXE — they go in COLLECT
+    exclude_binaries=True,  # ONEDIR mode
+    name='rfcp-server',
+    debug=False,
+    bootloader_ignore_signals=False,
+    strip=False,
+    upx=False,  # Don't compress — CUDA libs need fast loading
+    console=True,
+    icon=os.path.join(os.path.dirname(SPEC), 'rfcp.ico') if os.path.exists(os.path.join(os.path.dirname(SPEC), 'rfcp.ico')) else None,
+)
+
+coll = COLLECT(
+    exe,
+    a.binaries,
+    a.zipfiles,
+    a.datas,
+    strip=False,
+    upx=False,
+    upx_exclude=[],
+    name='rfcp-server',
+)
--- a/installer/rfcp.ico
+++ b/installer/rfcp.ico
--- a/installer/rthook_cuda_dlls.py
+++ b/installer/rthook_cuda_dlls.py
@@ -0,0 +1,24 @@
+# PyInstaller runtime hook for CUDA DLL loading
+# Must run BEFORE any CuPy import
+#
+# Problem: Windows Python 3.8+ requires os.add_dll_directory() for DLL search
+# PyInstaller ONEDIR mode puts DLLs in _internal/ which isn't in the search path
+
+import os
+import sys
+
+if sys.platform == 'win32' and getattr(sys, 'frozen', False):
+    # _MEIPASS points to _internal/ in ONEDIR mode
+    base = getattr(sys, '_MEIPASS', None)
+    if base and os.path.isdir(base):
+        os.add_dll_directory(base)
+        print(f"[CUDA DLL Hook] Added DLL directory: {base}")
+
+    # Also add CUDA_PATH if available (fallback to system CUDA)
+    cuda_path = os.environ.get('CUDA_PATH', '')
+    if cuda_path:
+        for subdir in ['bin', os.path.join('bin', 'x64')]:
+            d = os.path.join(cuda_path, subdir)
+            if os.path.isdir(d):
+                os.add_dll_directory(d)
+                print(f"[CUDA DLL Hook] Added CUDA_PATH: {d}")
--- a/rfcp-gpu-preflight.bat
+++ b/rfcp-gpu-preflight.bat
@@ -0,0 +1,64 @@
+@echo off
+echo ========================================
+echo  RFCP GPU Build — Pre-flight Check
+echo ========================================
+echo.
+
+echo [1] Python version:
+python --version
+
+echo.
+echo [2] CuPy status:
+python -c "import cupy; print(f'  CuPy {cupy.__version__}')"
+python -c "import cupy; d=cupy.cuda.Device(0); print(f'  Device: {d.id}'); print(f'  Memory: {d.mem_info[1]//1024//1024} MB')"
+
+echo.
+echo [3] CUDA runtime version:
+python -c "import cupy; v=cupy.cuda.runtime.runtimeGetVersion(); print(f'  CUDA Runtime: {v}')"
+
+echo.
+echo [4] CUDA_PATH environment:
+if defined CUDA_PATH (
+    echo   CUDA_PATH = %CUDA_PATH%
+) else (
+    echo   WARNING: CUDA_PATH not set!
+    echo.
+    echo   Checking common locations...
+    if exist "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA" (
+        for /d %%i in ("C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*") do (
+            echo   Found: %%i
+            echo.
+            echo   To fix, run:
+            echo     setx CUDA_PATH "%%i"
+            echo   Then restart terminal.
+        )
+    ) else (
+        echo   No CUDA Toolkit found in default location.
+        echo   CuPy bundles its own CUDA runtime, so this may be OK.
+        echo   But PyInstaller build might need it.
+    )
+)
+
+echo.
+echo [5] nvidia-smi:
+nvidia-smi --query-gpu=name,driver_version,memory.total --format=csv,noheader 2>nul
+if errorlevel 1 echo   nvidia-smi not found in PATH
+
+echo.
+echo [6] CuPy CUDA libs location:
+python -c "import cupy; import os; print(f'  {os.path.dirname(cupy.__file__)}')"
+python -c "import cupy._core.core" 2>nul && echo   cupy._core.core: OK || echo   cupy._core.core: FAILED
+
+echo.
+echo [7] fastrlock:
+python -c "import fastrlock; print(f'  fastrlock {fastrlock.__version__}')"
+
+echo.
+echo [8] PyInstaller:
+python -c "import PyInstaller; print(f'  PyInstaller {PyInstaller.__version__}')" 2>nul || echo   PyInstaller NOT installed! Run: pip install pyinstaller
+
+echo.
+echo ========================================
+echo  Pre-flight complete
+echo ========================================
+pause