v0.6.0

Breaking

• from_dataframe returns category_names as dicts: each encoded column's entry is now a {code: name} dict over every declared level (it was a list with one name per observed level). Names are keyed by the code they belong to, so a level the frame does not show cannot shift the others. Migration: names = schema.category_names[j] → names = list(schema.category_names[j].values()) for the names in code order, or look a code up directly with schema.category_names[j][code]. A Schema you write by hand may still use a list.
• NaN and infinity are rejected on numpy input: every constructor now raises a ValueError naming the first column that contains missing or infinite values. Before, such data was accepted and failed later — find_regions and any verb on the affected feature raised an unrelated error, importance returned NaN — while the effects of the clean columns still computed. effector has no missing-value semantics; impute or drop first (from_dataframe already rejected missing values).
• Integer-valued features print without decimals: a rule on a feature whose values are all integers (in display units) is written as the integers it admits — hr ≤ 6 where it read hr < 6.90, year = 0 where it read year = 0.00. This changes region labels in Partition.show(), plot titles and reports. A value that rounds to zero prints as 0.00, never -0.00.

Added

• select_regions(..., target=): an optional (N,) vector that replaces f̂(X) as the selection objective — e.g. the ground-truth labels, "which splits predict better" instead of "which splits explain the model". The curves still come from the effect; only the offsets fit and the R² denominator change.
• select_regions(..., relative_gain=): an optional hybrid gate. When set, a round's threshold is min(min_r2_gain, relative_gain * (1 - cum_r2)), so a near-saturated chain still accepts a split that explains at least relative_gain of the remaining unexplained variance. Never stricter than min_r2_gain; None (default) is the absolute rule, unchanged.
• select_regions(..., max_conditional_interactions=, max_partitions=): optional caps on the selected partitions — conditional interactions counted one per (feature, conditioning feature) pair, and the number of split features. A partition is accepted whole, so the sequence stops before the round that would exceed a cap: a capped chain is a prefix of the uncapped one. Splits a cap left out land in .skipped with the cap's name as the reason (interaction cap / partition cap in the printed ledger). CalmSequence stores and serializes the gate and the caps, only when set.
• from_dataframe(df, return_encoding=True) also returns a DataFrameEncoding, whose transform(df_new) encodes a later frame with the codes of the original one — columns matched by name, every level keeping its code, an unseen level raising.
• Schema.category_names accepts a {level value: name} dict per feature, next to the list form; levels absent from the data are fine.
• Partition.conditions(idx) — a region's rule as a list, one formatted condition per feature — and Partition.tree_lines() — the partition as a plain tree, one line per region, leaves with their number of instances.
• effector.ingestion.check_finite(data, feature_names) — the NaN / infinity check as a function.

Fixed

• A candidate partition with an empty region is skipped in select_regions (reason empty_region) instead of failing when its snapshot is built.
• Display units honor the schema's scale_y: every surface labeled "(target units)" — effector.plot_triage, CALM.plot_triage, and the report's importance/heterogeneity bars, triage plane, ranked tables, ledger heter columns, and heter_threshold chip — now bridges the model-unit scalars by scale_y["std"], so a model trained on a standardized target reads in the target's own scale (matching the curve plots, which already rescaled). Unbound reports render the same as bound ones: Report stamps scale_y/scale_x_list, and the unbound curve fallback applies them. The engine verbs (importance, heter_score) and all stamped/serialized payloads stay in model-output units, so find_regions thresholds and old dicts are unaffected.
• HTML report surfaces missed by the scale_y bridge: the decision-sequence table's heterogeneity column, the per-feature and global-baseline chips, the rejected-split note, and the per-leaf statistics (figure captions and the unbound leaf table) still rendered raw model-unit scalars while the rest of the page was bridged. They now use the same scale_y["std"] bridge and the ranked table's .4g format, so one page reads in one unit. The partition-tree <pre> dump keeps model units by design (it is the verbatim Partition.show() payload).