Release 0.34.0

[bbstats]

Version bump to 0.34.0 (pyproject.toml, chimeraboost/__init__.py) and the CHANGELOG's [Unreleased] renamed to [0.34.0] - 2026-09-26. Nothing else.

What's in 0.34.0

• refresh(X, y) and store_training_data (refresh(X, y): fold new rows into a fitted model (#131, slice 1) #170, issue Leaf refresh on new data: refresh(X, y) with opt-in stored training set #131 slice 1): fold new rows into a fitted regressor or binary classifier without a full retrain. Opt-in.
• The quantile head retrains its winner on all rows by default (refit_full, Quantile model: retrain the winner on all rows by default (refit_full) #180): called without an eval_set, it now learns from the rows it held out for early stopping. It improves CRPS on 35 of 36 Grinsztajn regression datasets (median +0.9%).
• Two more quantile candidates, "fixed" and "scaled" (Quantile model: fixed-width and scaled candidates in the default audition #177): the head now beats CatBoost MultiQuantile on 29 of 36 datasets.
• Linear-leaf fits are faster with identical results (refresh(X, y): fold new rows into a fitted model (#131, slice 1) #170).

What users will notice: quantile predictions change, and a quantile fit without an eval_set takes about 40% longer (the new candidates plus the retrain). audition=False, refit_full=False gives the single head as before. Ordinary regressors and classifiers produce the same outputs as 0.33.0: the exact-output snapshot is unchanged for every non-quantile configuration.

Checks on this branch

• Code identical to the tree that passed the full suite (1292 passed, 1 skipped); CI green on main's head.
• Built and checked: chimeraboost-0.34.0.tar.gz and chimeraboost-0.34.0-py3-none-any.whl pass twine check; the wheel carries all 15 source modules, including the new training_rows.py.

After you merge

I rebuild from main, upload to PyPI, tag v0.34.0 on the merge commit (pushing only the tag), create the GitHub release, and smoke-test pip install chimeraboost==0.34.0 in a throwaway environment.

🤖 Generated with Claude Code