Wire the guarded conversational RAG answer layer end-to-end

ingestion/pyproject.toml

@@ -3,10 +3,22 @@ name = "ingestion"
 version = "0.0.0"
 description = "Offline batch pipeline: PDF -> monographs -> chunks -> embeddings -> Qdrant"
 requires-python = ">=3.11"
-dependencies = ["pymupdf>=1.24", "numpy>=1.26", "scipy>=1.11"]
+dependencies = ["pymupdf>=1.24", "numpy>=1.26", "scipy>=1.11", "tiktoken>=0.7"]
 
 [project.optional-dependencies]
 dev = ["pytest>=7.4"]
+# Only the live probe and a real embedding run need these. The adapters and
+# their tests import neither, so the default install stays offline-capable.
+bedrock = ["boto3>=1.34"]
+local-embed = ["sentence-transformers>=3.0"]
+# `load/` talks to a VectorStore port; only `load.qdrant_repo` imports this,
+# lazily, so the load stage is tested in full with no server running.
+qdrant = ["qdrant-client>=1.7"]
+
+[tool.pytest.ini_options]
+markers = [
+    "integration: needs a live service (a local Qdrant); skips when absent",
+]
 
 [build-system]
 requires = ["setuptools>=68"]