[tgenman]

Summary

The CatBoost runtime currently returns class labels only, so clients cannot apply their own probability thresholds. Add predict_proba output selection while keeping predict as the default.

explicit request outputs > model parameters.extra.predict_fn > PREDICT_PROBA > predict

Clients can request predict, predict_proba, or both; PREDICT_PROBA=true sets the process-wide default. Probability outputs preserve class order and have shape [batch, classes], including for a single 1-D input.

Evidence

• Before: the runtime called CatBoostClassifier.predict() and returned class labels.
• After: runtimes/catboost/tests passes (10 tests), covering the existing default, environment flag, request output selection, precedence and invalid output.
• A local HTTP-v2 batch with a CatBoost classifier containing categorical and embedding features returned probabilities matching direct CatBoost inference (4 examples; max absolute difference 0).

Merge Danger

Door: Two-way; the default output remains predict when no option is set.

Blast Radius: CatBoost runtime. Explicit unsupported output names now return an inference error. PREDICT_PROBA is process-wide; the per-model setting or explicit request can override it.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[tgenman]

Hi @lc525 , could you please take a look at this PR or point me to the right maintainer for the CatBoost runtime? It adds an opt-in probability output while preserving the current default. Thanks!