feat: CUDA 13.2でgenerate_imageの実GPU生成に対応

CPU版PyTorchが選択されていた依存設定を修正し、RTX 5050上でStable Diffusion 1.5のMCP generate_imageを実行できる構成へ更新。併せてVAEとLoRAの仕様・実装・テストを明確化した。

CUDA / 依存管理:
- torch 2.14.0+cu132とtorchvision 0.29.0+cu132をruntime依存へ固定
- PyTorch CUDA 13.2のexplicit indexとtool.uv.sourcesをuv設定へ追加
- uv.lockをCUDA wheelに更新し、uv sync後もCPU版へ戻らないよう修正
- RTX 5050の認識、CUDA 13.2 runtime、GPU FP16演算を確認

モデル・オフライン実行:
- model.root配下のhuggingface cacheをDiffusersへ明示的に渡すよう変更
- checkpoint、外部VAE、ControlNetのsingle-file loadをlocal_files_onlyで統一
- SD 1.5 checkpointと公式pipeline構成・tokenizerによるオフラインloadを確認
- data/config.generate-image.tomlにText-to-Image用最小設定を追加
- data/MODELS.mdにモデル出典、容量、SHA-256、cache構成、検証結果を記録
- 4.27GBのcheckpoint、cache、生成画像はGit管理対象外を維持

VAE / LoRA:
- Stable Diffusion pipelineではVAE componentが必須であることを仕様・設計・設定文書に明記
- model.vaeはVAEの有無ではなく外部VAE差し替えの任意設定であることを明確化
- checkpoint内蔵VAEも外部VAEもない場合はpipeline初期化を失敗させる検証を追加
- 複数LoRAを起動時にロードし各weightでset_adaptersする既存動作を文書化
- generate_imageを含む全生成ツールへ設定済みLoRAが共通適用されることを明記
- 外部VAE差し替え、VAE欠落、複数LoRA適用のunit testを追加

動作確認:
- 実backendとin-process MCP Clientを使うsmoke_generate_image.pyを追加
- RTX 5050でSD 1.5、512x512、20 stepsのgenerate_imageを実行
- PNG ImageContent返却とseed、size、mode、timestamp、処理時間metadataを確認
- GPU生成時間18.313秒
- pytest: 32 passed
- Ruff lint/format: passed
- Pyright: 0 errors, 0 warnings
- uv lock check: passed

1 files changed, 99 insertions, 0 deletions

diff --git a/tests/test_pipeline_factory.py b/tests/test_pipeline_factory.py
new file mode 100644
index 0000000..454a5b1
--- /dev/null
+++ b/tests/test_pipeline_factory.py
@@ -0,0 +1,99 @@
+from __future__ import annotations
+
+from dataclasses import replace
+from types import SimpleNamespace
+
+import pytest
+
+from mcp_sd15.config import LoRAConfig
+from mcp_sd15.errors import ConfigError
+from mcp_sd15.pipeline_factory import PipelineFactory
+
+
+class FakePipeline:
+    def __init__(self, *, with_vae: bool = True) -> None:
+        self.vae = object() if with_vae else None
+        self.scheduler = SimpleNamespace(config={})
+        self.lora_loads: list[dict[str, object]] = []
+        self.adapters: tuple[list[str], list[float]] | None = None
+        self.device: str | None = None
+
+    def load_lora_weights(self, directory: str, **kwargs: object) -> None:
+        self.lora_loads.append({"directory": directory, **kwargs})
+
+    def set_adapters(self, names: list[str], *, adapter_weights: list[float]) -> None:
+        self.adapters = (names, adapter_weights)
+
+    def set_progress_bar_config(self, *, disable: bool) -> None:
+        assert disable is True
+
+    def to(self, device: str) -> FakePipeline:
+        self.device = device
+        return self
+
+
+class FakeScheduler:
+    @classmethod
+    def from_config(cls, config: dict[str, object]) -> object:
+        return {"scheduler": cls.__name__, **config}
+
+
+def fake_diffusers(pipeline: FakePipeline, external_vae: object) -> SimpleNamespace:
+    class StableDiffusionPipeline:
+        @classmethod
+        def from_single_file(cls, path: str, **kwargs: object) -> FakePipeline:
+            pipeline.base_load = {"path": path, **kwargs}  # type: ignore[attr-defined]
+            return pipeline
+
+    class AutoencoderKL:
+        @classmethod
+        def from_single_file(cls, path: str, **kwargs: object) -> object:
+            pipeline.vae_load = {"path": path, **kwargs}  # type: ignore[attr-defined]
+            return external_vae
+
+    return SimpleNamespace(
+        StableDiffusionPipeline=StableDiffusionPipeline,
+        AutoencoderKL=AutoencoderKL,
+        DPMSolverMultistepScheduler=FakeScheduler,
+        DDIMScheduler=FakeScheduler,
+        EulerDiscreteScheduler=FakeScheduler,
+        EulerAncestralDiscreteScheduler=FakeScheduler,
+    )
+
+
+def test_base_pipeline_requires_vae_and_applies_all_loras(app_config, monkeypatch) -> None:
+    model = replace(
+        app_config.model,
+        vae="external.safetensors",
+        loras=(
+            LoRAConfig("detail.safetensors", 0.8),
+            LoRAConfig("style.safetensors", 0.5),
+        ),
+    )
+    factory = PipelineFactory(replace(app_config, model=model))
+    pipeline = FakePipeline()
+    external_vae = object()
+    runtime = (SimpleNamespace(float16="float16"), fake_diffusers(pipeline, external_vae))
+    monkeypatch.setattr(factory, "_runtime", lambda: runtime)
+
+    result = factory.create_base()
+
+    assert result is pipeline
+    assert pipeline.vae is external_vae
+    assert pipeline.adapters == (["configured_lora_0", "configured_lora_1"], [0.8, 0.5])
+    assert [item["weight_name"] for item in pipeline.lora_loads] == [
+        "detail.safetensors",
+        "style.safetensors",
+    ]
+    assert all(item["local_files_only"] is True for item in pipeline.lora_loads)
+    assert pipeline.device == "cuda"
+
+
+def test_checkpoint_without_vae_fails_pipeline_initialization(app_config, monkeypatch) -> None:
+    factory = PipelineFactory(app_config)
+    pipeline = FakePipeline(with_vae=False)
+    runtime = (SimpleNamespace(float16="float16"), fake_diffusers(pipeline, object()))
+    monkeypatch.setattr(factory, "_runtime", lambda: runtime)
+
+    with pytest.raises(ConfigError, match="does not provide a VAE"):
+        factory.create_base()