#1506 [Windows/NVIDIA 0.1.40.2] CPU vision crashes before READY (0xC0000005), GPU vision works

Description

## Summary

CPU vision fails to start on Windows with the prebuilt Strata 0.1.40.2 NVIDIA engine.

GPU vision works correctly with the exact same Strata installation, model and mmproj.

When vision is configured on the CPU, `strata-vision.exe` exits before printing `READY`, and the Strata server raises:

```text
loading the vision encoder ...
Traceback (most recent call last):
  File "C:\Users\Montage1\Documents\Strata\serve\server.py", line 5611, in <module>
    sys.exit(main())
  File "C:\Users\Montage1\Documents\Strata\serve\server.py", line 5411, in main
    vision = Vision(...)
  File "C:\Users\Montage1\Documents\Strata\serve\server.py", line 1741, in __init__
    self._start()
  File "C:\Users\Montage1\Documents\Strata\serve\server.py", line 1763, in _start
    raise RuntimeError("the vision encoder did not start: " + line.strip())
RuntimeError: the vision encoder did not start:
```

## Environment

- OS: Windows 10 Pro 22H2
- OS build: 19045
- Strata git commit: `e8ca9afd03d839d4f8dbbe82dffce7f8a3bafd7a`
- Strata engine: 0.1.40.2
- Engine source: release
- CUDA runtime reported by BUILD.json: CUDA 13.0

### Hardware

- GPU: NVIDIA GeForce RTX 3080 10 GB
- NVIDIA driver: 581.08
- GPU compute capability: 8.6
- CPU: Intel Core i9-11900 (AVX-512)
- RAM: 64 GB
- PCIe: 4.0 x16

## Model

- Family: Swift 1.5
- Quant: IQ3_XXS
- Context: 98304
- KV cache: int8
- KV resident: 32768
- Vision mmproj: `mmproj-Swift-Qwen3.8-Flash-Next-BF16.gguf`

## Setup command

```bat
START-HERE.bat --setup --yes --family swift --model IQ3_XXS --context 98304 --vision cpu --no-start
```

Setup completes normally and reports:

```text
[ok] images: on (encoder on the CPU)
[ok] ready-made engine already installed
[ok] NVIDIA CUDA libraries (cuBLAS, CUDA runtime; ~0.4 GB) already installed
[ok] vision encoder already downloaded
[ok] model prepared
[ok] MTP draft layer
```

The generated vision config contains:

```text
exe        : C:\Users\Montage1\Documents\Strata\engine\strata-vision.exe
gpu        : False
max_tokens : 300
threads    : 8
```

## Reproduction

Starting Strata afterwards:

```bat
START-HERE.bat --port 8081
```

fails immediately at:

```text
loading the vision encoder ...

RuntimeError: the vision encoder did not start:
```

No error text from `strata-vision.exe` is forwarded before the process exits.

## Direct `strata-vision.exe` test

Running the helper directly initially exits silently with:

```text
-1073741515
```

which corresponds to Windows status:

```text
0xC0000135
```

(`STATUS_DLL_NOT_FOUND`)

The configured Strata CUDA library directory is:

```text
C:\Users\Montage1\Documents\Strata\.venv\Lib\site-packages\nvidia\cu13\bin\x86_64
```

It contains:

```text
cublas64_13.dll
cublasLt64_13.dll
cudart64_13.dll
```

The following MSVC runtime DLLs are also present in `System32`:

```text
VCRUNTIME140.dll
VCRUNTIME140_1.dll
MSVCP140.dll
```

After explicitly prepending Strata's CUDA library directory to PATH:

```powershell
$c = Get-Content .\strata-swift-iq3_xxs.json -Raw | ConvertFrom-Json

$oldPath = $env:PATH
$dirs = @($c.lib_dirs | Where-Object { $_ -and (Test-Path $_) })

if ($dirs.Count) {
    $env:PATH = (($dirs -join ";") + ";" + $env:PATH)
}

& $c.vision.exe `
    --mmproj $c.vision.mmproj `
    --model $c.vision.model `
    --threads $c.vision.threads `
    --max-tokens $c.vision.max_tokens

$LASTEXITCODE

$env:PATH = $oldPath
```

`strata-vision.exe` still prints nothing, but the exit code changes to:

```text
-1073741819
```

which corresponds to:

```text
0xC0000005
```

(access violation)

So after supplying the CUDA DLL directory, the executable gets further than the DLL-loading failure but appears to crash in the CPU vision path.

## GPU vision works

The same install, model and mmproj work correctly when vision is configured on the GPU.

With GPU vision:

```text
strata-vision: model files loaded in 1.1 s
strata-vision: warmed up at 1024 image tokens in 0.5 s
```

and the model starts and processes images correctly.

Therefore the issue seems specific to CPU vision / the Windows prebuilt vision helper rather than the model or mmproj.

## Why CPU vision matters on this hardware

On this 10 GB RTX 3080, GPU vision has a large effect on the MoE expert cache.

Without GPU vision:

```text
~1371 resident experts
~2.26 GiB expert cache
```

With GPU vision:

```text
530 resident experts
~0.88 GiB expert cache
```

So CPU vision would be particularly valuable on lower-VRAM NVIDIA cards.

For reference, the model itself otherwise runs very well on this machine:

- Swift 1.5 IQ3_XXS
- 98304 context
- ~26–33 tok/s decode in normal Open WebUI use

Happy to test a different `strata-vision.exe` build, collect additional logs, or run any diagnostic command that would help.