…roxy

The similarity search API selected entities by a CURIE prefix standing in for
"things of some category in some species". That only works while prefix,
category and taxon coincide one-to-one, and they do not: mouse models are split
across the MGI and MMRRC prefixes so no group can name them, while MGI alone
covers mouse genes, genotypes and variants at once.

Search now takes explicit `categories` / `taxa` / `prefixes`, which AND
together. GET /semsim/filters enumerates the valid combinations from the KG
rather than a hand-maintained enum, and POST /search takes a `filter` object.
SemsimSearchGroup keeps working: it expands to the explicit filters it always
implicitly meant.

Enabling changes:

- DEFAULT_ASSOCIATIONS selects by predicate rather than by association
  category. The old gene+disease allowlist structurally hid all 589,030
  GenotypeToPhenotypicFeature edges -- every MGI and MMRRC mouse model -- plus
  variants and cases, with no way for a caller to opt in. Because the pool is
  now heterogeneous, the legacy groups must state their category to keep their
  old meaning; verified they still return a single category and the same top
  hits (MECP2/Mecp2/Rett syndrome for a Rett query).

- _termset_search scores at the phenotype level instead of per (entity,
  phenotype). A phenotype's similarity to a query term does not depend on which
  entity carries it, so the old query re-expanded each phenotype's ancestors
  once per annotated entity: across mouse genotypes, 447,110 association rows
  covering 11,530 distinct MP terms, ~39x redundant. Scoring all 81,151 mouse
  genotypes against Rett syndrome goes from 288.7s to 2.3s (~125x). Verified
  equivalent, not assumed: both paths return all 81,151 entities with a maximum
  absolute score difference of 2.0e-15. Exact because `closure` holds no
  duplicate (subject, object) pairs, so count(*) and count(DISTINCT a) agree.

- _flat computes closure sizes for entities the baked `closure_size` table
  omits. It covers genes and diseases but no genotypes, and _flat inner-joined
  it, so hybrid_search -- the production default -- returned an empty result
  set for every mouse model with no error at all.

Filtering on a `nodes` table lacking category/in_taxon now raises instead of
matching nothing, so a filter that cannot be honored never looks like a
legitimately empty result. Same reasoning for SemsimianService, which can only
express a single prefix and now refuses anything else rather than silently
answering with an unfiltered superset.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PTYEKvvmBKgaK3893Kxog2