Release 0.9.0

[orgoca]

What 0.9.0 contains

PR	Change
#77	compile_selected, the three-partition protocol, the frozen ceiling with K refits, the NumPy batch scorer, provenance, check 11, sweep warnings (closes #75)
#78	The benchmark through compile_selected, the language migration, benchmarks/sync_docs.py

Why a minor release

compile_selected, compileml.batch.score_batch, train_whitebox(backend=) and validation check 11 are new public API. git diff --stat v0.8.0..HEAD over runtime/, export/ and docs/ARTIFACT_SPEC.md is empty apart from one wording fix in the spec's aside.

This PR

• __version__ → 0.9.0; CITATION.cff version and date-released; the [0.9.0] changelog heading with a summary.

• The benchmark's latency is now measured in a fresh process after a 30 s rest. In Run the benchmark through compile_selected and stop describing distillation as the method #78 every millisecond figure had come out 1.3–2.3× slower than the same artifact measured on its own: selection refits the ceiling with a multithreaded booster and compiles forty candidates, and timings taken in the same process straight afterwards suffered from lingering worker threads and a hot CPU. An isolated probe of the committed artifact read 0.39–0.41 ms where the in-process figure was 0.94.

  Measured properly:

  	Run the benchmark through compile_selected and stop describing distillation as the method #78 (in-process, throttled)	this PR (fresh process)	0.5.1, 120 trees
  score + band + PD, median	0.034 ms	0.021 ms	0.030 ms
  full explained decision, median	0.696 ms	0.416 ms	0.625 ms
  cost sweep, per tree at p = 23	0.889 ms	0.692 ms	0.705 ms

  The sweep's models are identical to 0.5.1's, with identical tree-walk counts; the per-tree timings now land within 3% of that run. Retention is unchanged, as it must be: same configuration selected, same hash. results.json, the README table, the attribution page and the FAQ are regenerated by benchmarks/sync_docs.py, and the changelog's latency note now describes the fix rather than the erratic machine.

• The sweep_whitebox examples in the tuning guide and README pass alpha_grid explicitly, so copied code does not trigger the new warning. examples/README.md no longer counts the validation checks.

ruff, black, full pytest and mkdocs build --strict pass.

After merge

Wait for the matrix on the merge commit to go green, then tag that commit:

git checkout main
git pull
git show HEAD:src/compileml/__init__.py | grep __version__   # must print 0.9.0
git tag v0.9.0
git push origin v0.9.0

Then publish the GitHub release from v0.9.0 and approve the pypi gate.