P1: host-agnostic quercus-core with agent loop, providers, and tool schemas

- Core types: Rational time, EntityId, ActionBatch, PngBytes (no Pillow dep)
- Config: ~/.quercus/config.toml with 0600 perms, env var override for CI
- LLMProvider: Claude (anthropic SDK) and OpenAI-compatible endpoint
  (covers custom gateways and llama.cpp server mode); graceful degradation
  without API keys
- Tool schemas: 26 curated tools with JSON Schema and read_only flags
- AgentLoop: sync dialogue loop with confirmation gate (read-only batches
  bypass, rejected mutations are reported back to the LLM), PNG frame
  feedback for multimodal providers
- Session log: JSONL recording with frame sha256 hashes, replayable
- HostAdapter ABC (signatures only, aligned with OPP/1 method families)
- Mock LLM provider (scripted/replay/record) for network-free CI

Tests: 69 passed via uv run pytest (no network access)

pyproject.toml

@@ -0,0 +1,27 @@
+[project]
+name = "quercus"
+version = "0.1.0"
+description = "Quercus AI 剪辑助手 — monorepo（workspace 根，虚拟包）"
+requires-python = ">=3.11"
+dependencies = [
+    "quercus-core",
+]
+
+[dependency-groups]
+dev = [
+    "pytest>=8.0",
+]
+
+[tool.uv]
+# 根目录不是可安装包，仅作 workspace 锚点
+package = false
+
+[tool.uv.sources]
+quercus-core = { workspace = true }
+
+[tool.uv.workspace]
+members = ["core"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+addopts = "-q"