v0.2.4

Changed

• Engine floor moves to kglite 0.16.2 — edge properties now carry their real
  types. One 0.16.2 fix reaches us, and it reaches everything: bulk-loaded
  edge properties (add_connections, the only edge path codingest uses)
  recorded every property as Unknown instead of its observed type. Measured
  before/after on the same corpus, CALL db.schema.relTypeProperties() moves
  :CALLS.call_count from ["Unknown"] to ["Long"] and
  :CALLS.import_backed to ["Boolean"], so schema consumers — graph_overview,
  Neo4j-shaped clients reading db.schema.* / apoc.meta.* — get a typed
  edge schema for the first time. A graph built before this keeps its
  Unknowns until its edges are rewritten: rebuild stale .kgl artifacts,
  re-reading them on the new engine does not fix them. The property values
  are unchanged, so the frozen parity goldens are byte-identical across the
  bump. Nothing else in 0.16.2 touches us (Bolt-server Neo4j/APOC
  compatibility, Java bindings, an id(n) projection fast path).

Fixed

• Published crates no longer carry local working state. The dev-docs/
  ignore rules are root-anchored (they must be, for the
  !/dev-docs/bench/scripts/ re-include), so a dev-docs/ created under a
  crate directory was merely untracked — and cargo package ships
  untracked-but-unignored files. A stray crates/codingest/dev-docs/ had put
  15 scratch files into the codingest package. Caught before publish;
  0.2.3 on crates.io is clean. Fixed at both ends: crates/**/dev-docs/ is
  ignored, and a release gate now asserts cargo package --list for all four
  publishable crates carries no dev-docs/ or inbox/ path — read from
  cargo's real list, not re-derived from the ignore rules.