Narwhalify model layer

[AdrianSosic]

Extends the narwhalification series to the surrogates, GP component factories, acquisition functions, and the Bayesian recommender layer.

[Copilot]

🟡 Changes recommended

A confirmed regression in Surrogate.posterior_stats uses vals.numpy() instead of a safe CPU conversion, which can raise at runtime when tensors are not on CPU.

Once you've addressed the issues Copilot identified, you can request another Copilot review.

Pull request overview

This PR continues the “narwhalification” effort by updating the modeling stack (objectives → surrogates → acquisition functions → Bayesian recommenders/campaign APIs) to accept/return Narwhals-compatible dataframe/series inputs, reducing direct pandas coupling in the model layer.

Changes:

• Switch many model-layer interfaces from pd.DataFrame to Narwhals typing (IntoDataFrame, IntoDataFrameT, IntoSeries) and use narwhals.stable.v2 conversions internally.
• Refactor objective missing-value handling to a private, backend-preserving helper (Objective._handle_missing_values).
• Update surrogate posterior statistics and acquisition evaluation paths to return backend-native objects with index preservation helpers.

File summaries

File	Description
CHANGELOG.md	Document breaking change: Objective.handle_missing_values made private.
baybe/surrogates/gaussian_process/presets/hvarfner.py	Narwhals-typed measurements for GP component factories.
baybe/surrogates/gaussian_process/presets/edbo.py	Narwhals-typed measurements for kernel/likelihood factories.
baybe/surrogates/gaussian_process/presets/edbo_smoothed.py	Narwhals-typed measurements for smoothed preset factories.
baybe/surrogates/gaussian_process/presets/chen.py	Narwhals-typed measurements for Chen kernel factory.
baybe/surrogates/gaussian_process/presets/botorch.py	Narwhals-typed measurements for BoTorch preset kernel factory.
baybe/surrogates/gaussian_process/presets/baybe.py	Narwhals-typed measurements across BayBE preset factories.
baybe/surrogates/gaussian_process/core.py	Store GP model context measurements as nw.DataFrame.
baybe/surrogates/gaussian_process/components/mean.py	Narwhals-typed measurements for mean factories.
baybe/surrogates/gaussian_process/components/likelihood.py	Narwhals-typed measurements for likelihood factories.
baybe/surrogates/gaussian_process/components/kernel.py	Narwhals-typed measurements across kernel factory protocols/implementations.
baybe/surrogates/gaussian_process/components/generic.py	Update GP component factory protocol to accept Narwhals-compatible dataframes.
baybe/surrogates/gaussian_process/components/fit_criterion.py	Narwhals-typed measurements in fit-criterion factory.
baybe/surrogates/composite.py	Use Narwhals conversions during per-quantity fitting and horizontal concat of posterior stats.
baybe/surrogates/base.py	Accept Narwhals inputs for fit/posterior APIs; return backend-native posterior stats with index copying.
baybe/recommenders/pure/bayesian/base.py	Pass Narwhals-compatible frames through Bayesian recommender training/acqf setup and acquisition evaluation.
baybe/recommenders/naive.py	Stop forcing pandas conversion when delegating to Bayesian recommender acqf setup.
baybe/objectives/base.py	Introduce private, backend-preserving _handle_missing_values; update dataframe utilities usage.
baybe/campaign.py	Widen campaign acquisition/pending-experiments typing to Narwhals-compatible frames/series.
baybe/acquisition/base.py	Return backend-native series from acquisition evaluation with index copying.
baybe/acquisition/acqfs.py	Return integration points in the configured default dataframe backend via Narwhals conversion.
baybe/acquisition/_builder.py	Make builder generic over dataframe backend and preserve backend in target-config selection.

Review details

• Files reviewed: 22/22 changed files
• Comments generated: 1
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.