feat: Phase 1+2 — project setup and core infrastructure

- requirements.txt, config.yaml, .env, Dockerfile, docker-compose.yml
- app/core: config (YAML+env override), logging (JSON structured),
  exceptions (typed hierarchy), json_utils (Markdown fence stripping)
- app/clients: LLMClient ABC + ZhipuAIClient (run_in_executor),
  StorageClient ABC + RustFSClient (boto3 head_object for size check)
- app/main.py: FastAPI app with health endpoint and router registration
- app/core/dependencies.py: lru_cache singleton factories
- tests/conftest.py: mock_llm, mock_storage, test_app, client fixtures
- pytest.ini: asyncio_mode=auto
- 11 unit tests passing

config.yaml

@@ -0,0 +1,19 @@
+server:
+  port: 8000
+  log_level: INFO
+
+storage:
+  buckets:
+    source_data: "source-data"
+    finetune_export: "finetune-export"
+
+backend: {}   # callback_url injected via BACKEND_CALLBACK_URL env var
+
+video:
+  frame_sample_count: 8    # uniform frames sampled for video-to-text
+  max_file_size_mb: 200    # video size limit (override with MAX_VIDEO_SIZE_MB)
+  keyframe_diff_threshold: 30.0  # grayscale mean-diff threshold for keyframe detection
+
+models:
+  default_text: "glm-4-flash"
+  default_vision: "glm-4v-flash"