v0.2.2

Fixed

• The Python-side kglite floor is back in lockstep with the Rust one. 0.2.1
  moved the Rust engine floor to kglite 0.15.13 but left every other
  declaration of that floor at 0.15.11, so the published 0.2.1 wheel requires
  kglite>=0.15.11,<0.16 while the Rust engine compiled into it is 0.15.13.
  The wheel's Rust kglite writes the .kgl bytes and the separately-installed
  Python kglite reads them, which is exactly the skew
  test_kglite_cargo_and_python_floors_are_in_lockstep exists to prevent — it
  went red in CI on the 0.2.1 commit and the tag was pushed before that run
  reported. In practice pip resolves the range to 0.15.13 anyway, so the
  exposure is limited to an install that deliberately pins 0.15.11; it is
  corrected here rather than left as a latent mismatch.

  The floor turned out to be declared in 15 places across 8 files, not the
  one the test named: pyproject.toml (the requirement and the comment
  asserting the lockstep rule), .github/workflows/ci.yml (the pinned
  kglite==, which is what actually decides the engine CI exercises — the same
  site that was left behind at 0.15.5 once before), the pip install line in
  codingest-py's import-failure message, the floor rationale in the root and
  codingest manifests, and seven prose declarations across README.md,
  docs/index.md, docs/migrating-from-kglite-code-tree.md and
  docs/python-api.md. A further 7 mentions were deliberately left at
  0.15.11: they cite when an API first appeared (rev.rs's
  "set_node_property (kglite >= 0.15.11)", docs/index.md's "0.15.11 made
  reachable", the floor history in Cargo.toml) and remain true. A version
  declaration must move; a version citation in a historical statement must
  not.

Known issues

• test_committed_baseline_is_well_formed[0.2.0] is red and is not silenced
  here. The committed 0.2.0 perf baseline records a control query at 0.012 ms
  against its own 0.013 ms noise floor, so the control cannot resolve what it is
  meant to police. This predates 0.2.1 — it was already red on the 0.2.0 release
  commit — and it shares a root cause with the VOID recorded in [0.2.1]: the
  control is anchored_callers, which kglite 0.15.13 then made 5.7x faster
  still. Both available quick fixes (raise the recorded floor, or drop the
  baseline) are the re-baseline-to-silence move this project forbids at the
  parity gate and forbids here for the same reason. Choosing a control the
  engine cannot reach is a design change and is tracked as such; no new baseline
  was captured for 0.2.1 or 0.2.2, because a capture taken now would commit an
  artifact its own well-formedness test rejects.