Qdrant supplies candidate ids and scores and nothing else. Mind::rank drops
every hit the mind does not hold, then lets CultNet's selection decide
membership over the whole mind with its keys narrowed to the candidates, so
in_force and cited mean what they mean in query. Order is score then ordinal;
the page is cut to the selection's limit and top_k; next is never set. A
cursor with semantic, blank text and an out-of-range top_k are refused by name
before anything is embedded. Mind::query loses its semantic parameter and the
"index is not wired" refusal; the published request schema's description of
SemanticQuery follows.
The serve loop never waits on the embedder: a semantic query becomes a Search
held beside its session, the worker embeds (behind the model's query
instruction) and searches between batches, and the loop answers when the
worker has or when ServeOptions::search_timeout passes. A worker that is
Failing or Refused refuses at once, typed Unavailable naming its own error.
VectorIndex gains search; Qdrant's adapter sends no filter.
Model identity is re-read before every flush and before every search: a model
re-pulled under the same dimensions but a new digest forgets the
reconciliation instead of mixing vectors, and the next step rebuilds. An idle
worker forgets the reconciliation every Backoff::recheck (60 s), so a lost
collection is found without a write. Both go through Projector::reconcile,
still the one judge.
Also fixes a race in the worker loop found by the mutation runs: entries were
taken from the inbox under one lock and published as wanted under another, so
whoami could read a stale Current with nothing queued while the worker held
them. Taking, absorbing and publishing are now one critical section.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
5eea49f756