feat: add Qdrant as alternative storage backend

[RobertoGEMartin]

Summary

Adds Qdrant as an opt-in alternative to ChromaDB for the palace vector store. This resolves scalability issues when palaces exceed ~100K drawers (related to #688).

• New QdrantBackend + QdrantCollection in mempalace/backends/qdrant.py, implementing BaseCollection
• Backend selection via config.json: {"backend": "qdrant"} (ChromaDB remains the default)
• Full ChromaDB filter compatibility: $eq, $ne, $in, $nin, $and, $or, $contains
• Deterministic UUID5 ID mapping for reproducible migrations
• Lazy imports — qdrant-client and sentence-transformers only loaded when Qdrant backend is active

Why Qdrant?

Metric	ChromaDB	Qdrant
Max tested drawers	~100K (SQLite variable limit)	168K+ (production-tested)
Storage (168K drawers)	~3.4 GB	~800 MB (~4x smaller)
Filter queries at scale	Degrades (col.get() fails)	Stable (payload indexes)

Installation

pip install mempalace[qdrant]

Then in ~/.mempalace/config.json:

{"backend": "qdrant"}

Changes

File	Change
mempalace/backends/qdrant.py	New: QdrantBackend + QdrantCollection (411 lines)
mempalace/backends/__init__.py	Conditional Qdrant exports
mempalace/palace.py	Config-driven backend selection
mempalace/config.py	Added backend property
pyproject.toml	Added qdrant optional dependency
tests/test_backends_qdrant.py	20 tests (unit + integration)
README.md	Qdrant backend documentation

Test plan

• All 20 Qdrant tests pass (17 unit + 3 integration)
• All 6 existing ChromaDB backend tests still pass (zero regressions)
• Ruff lint passes on all changed files
• Production-tested with 168K+ drawers across 8 wings
• CI pipeline validation

🤖 Generated with Claude Code