Feature/qdrant ltm foundation

[matstech]

Qdrant LTM Backend

Summary

This PR adds Qdrant as an optional long-term memory vector backend alongside the existing Chroma implementation.

The main goal is to make DMF backend-neutral at the LTM boundary without changing the default runtime behavior: Chroma remains the default, while Qdrant can be enabled explicitly with storage_type = "qdrant" and qdrant_mode = "memory".

Why

DMF previously coupled long-term memory persistence and vector search to Chroma-specific APIs. That made adding another vector store risky because filtering, raw-record codecs, card lookup, and hook construction were spread across backend-specific assumptions.

This implementation introduces Qdrant through the same public LTM contract used by Chroma, keeping raw records canonical and preserving the existing card-to-raw resolution semantics. Qdrant is intentionally limited to Local Mode in this PR so the integration can be validated without introducing server lifecycle, auth, or persistence concerns.

What Changed

• Added backend-neutral LTM primitives:

  • shared raw/card payload codecs;
  • shared vector hit types;
  • RecallFilter for role, record id, interaction range, timestamp range, and card kind filtering;
  • runtime factory wiring for Chroma, Qdrant, File, and Null hooks.

• Added Qdrant Local Mode:

  • optional qdrant-client dependency via the qdrant extra;
  • in-memory QdrantClient(":memory:") connection mode;
  • deterministic UUIDv5 point IDs for raw records and cards;
  • COSINE scoring mapped back to DMF distance/threshold semantics;
  • raw search, card search, count, read-all, clear, and collection validation.

• Preserved existing Chroma behavior:

  • default config remains Chroma embedded;
  • legacy import shims remain available;
  • Chroma card filtering still resolves card hits through source raw records.

• Extended benchmark tooling:

  • benchmark runner can select --backend chroma|qdrant;
  • Qdrant benchmark does not require Docker;
  • final console summary is computed from the produced JSON report;
  • summary now separates completed-case retrieval rate from overall dataset retrieval rate;
  • Ollama requests disable model thinking to avoid empty message.content with reasoning models.

• Updated docs for Qdrant install, config, Local Mode limits, raw/card collections, score conversion, filters, errors, and API references.

Test Coverage

• Unit coverage for Qdrant client setup, optional dependency behavior, filters, codecs, hook operations, cards, malformed payloads, collection validation, and concurrency basics.
• Backend contract tests comparing Chroma and Qdrant behavior for raw archival, recall, thresholding, count/read/clear, and framework-level recall flows.
• Regression coverage for Chroma, File, Null, factory wiring, legacy imports, config parsing, benchmark utilities, and docs build.

Validation run:

poetry check --lock
make compile
make test
make docs-build
make build
git diff --check
make chroma-up
make test-integration
make chroma-down

Results recorded during final gate:

• make test: 745 passed, 1 deselected
• Chroma integration test: 1 passed
• docs and wheel build: passed

[ezimuel]

LGTM, thanks!