fix: make aiana installable and compatible with qdrant-client >= 1.12

[ry-ops]

Summary

Three fixes so aiana's full (Qdrant-backed semantic) setup installs and actually returns results. Verified end-to-end on a fully local stack (local Qdrant + local sentence-transformers embedder + a local MLX model via OpenCode).

Bug 1 — pyproject.toml is unparseable (pip install fails)

Two optional-dependency entries had malformed version specifiers (doubled operator + stray quote):

"pytest>=">=9.0.3"          # -> "pytest>=9.0.3"
"cryptography>=">=46.0.6"   # -> "cryptography>=46.0.6"

Invalid TOML (tomllib.TOMLDecodeError: Unclosed array (at line 59)), so the package cannot be installed at all.

Bug 2 — QdrantStorage.search() calls a removed client method

qdrant-client >= 1.12 removed QdrantClient.search() in favor of QdrantClient.query_points(). The old call raised AttributeError, which _memory_search()'s try/except silently swallowed — so semantic search just returned nothing.

results = self.client.query_points(
    collection_name=COLLECTION_NAME, query=query_vector, limit=limit,
    query_filter=search_filter, score_threshold=min_score, with_payload=True,
).points

Bug 3 — default min_score too strict (empty results for relevant queries)

With all-MiniLM-L6-v2, clearly-relevant memories score ~0.40–0.48 for natural-language queries, while irrelevant content scores < 0.1. Measured example:

query	top match	score
"what model does the MLX lab use"	"The MLX lab runs Qwen3.5-4B-4bit…"	0.475
"MLX lab model"	(same)	0.404

The old min_score=0.5 floor dropped these relevant top hits, so memory_search returned nothing for reasonable queries. Lowered the default to 0.3, which surfaces the correct top result while still filtering noise.

Verification

Python 3.13, qdrant-client 1.18.0, local Qdrant + all-MiniLM-L6-v2:

• pip install -e . succeeds.
• memory_add stores a 384-dim vector in aiana_memories.
• memory_search retrieves it via query_points and returns relevant hits at the new threshold.
• Driven live: a local MLX model (Qwen3.5-4B) via OpenCode calls aiana_memory_add then aiana_memory_search; both requests are visible in the Qdrant access log (PUT .../points, POST .../points/query).
• Backends init: SQLite + Redis + Qdrant + embedder (mem0 cleanly falls back with no LLM key configured).

🤖 Generated with Claude Code