Qdrant sink

When

kind: vector_vocab + backend: qdrant stores embeddings.
Not for kind: registry. Observation history stays in PostgreSQL.

DRTML

datasets:
  lemma_vectors:
    role: output
    backend: qdrant
    kind: vector_vocab
    uniq: lemma_token_id
    columns:
      lemma_token_id: string
      kind: string
      vector: list<float32>
    qdrant:
      connection: qdrant_app
      collection: lemma_vectors_v1
      vector_size: 128
      distance: Cosine
      vector_column: vector
      point_id_column: lemma_token_id

Rules

1. Dual-write with parquet is forbidden.
2. flush.boundary_columns → error.
3. Geometry (hubness, L2) is computed on the driver/post-train path, not by partition workers.
4. Registry entries may declare qdrant_metric capability names such as vector_l2_mean, but the generic dataset-metric dispatcher does not currently consume that field. Live Qdrant geometry uses the dedicated embed-train / Qdrant geometry modules.

Transport: db/qdrant/. Adapter in backend_builtins.