feat(model): tree/tabular backend + SHAP TreeExplainer end-to-end (#246)

[stanlrt]

Summary

Closes #246. Adds a tree/tabular model backend so SHAP TreeExplainer runs end-to-end. XGBoost is the MVP runtime; the architecture is a shared TabularTreeBackend base plus per-runtime subclasses, so sklearn (#295) and LightGBM (#296) land as ~15-line subclasses with no seam change. Additive only — new backend, new tree extra, a defaulted ForwardOutput.output_kind field, and a keyword-only prediction_summaries arg. No breaking changes.

Acceptance criteria → where

Criterion	Where
TreeExplainer runs end-to-end on a real tree model via a documented backend	test_tree_explainer_end_to_end_xgboost; docs in bd3e90e
requires reflects true capability (TREE_MODEL, not AUTOGRAD placeholder)	4e95f22
Non-tree backend → clear BackendIncompatibilityError (not SHAP-internal)	test_tree_explainer_gates_out_on_torch_backend; error hint in 4e95f22
Predictions/metrics behave sensibly for tree outputs	OutputKind + softmax skip (c0050c2, 5d2adb3)
Docs reflect supported methods + new backend	bd3e90e

How it works

• EstimatorProvider protocol + TREE_MODEL shape-capability binding hands the raw fitted estimator to TreeSHAP through the Make the model-module explanation contract backend-agnostic (drop the nn.Module bridge) #209 agnostic contract — no nn.Module facade (f2c1f4d).
• TabularTreeBackend (e07bfde) owns the torch→numpy→predict_proba→torch bridge, fitted_estimator(), CPU/classification defaults. XGBoostBackend (586c594) is a concrete subclass loading .ubj, registered provides={TREE_MODEL, PREDICT_PROBA}.
• Tree backends emit probabilities, not logits. ForwardOutput.output_kind is stamped from PREDICT_PROBA (c0050c2); classification prediction_summaries skips softmax for probabilities (5d2adb3). Affects confidence display only — metrics argmax regardless, so metric numbers are unchanged.
• predict_callable (inherited) exposes the probability bridge, so model-agnostic SHAP explainers (e.g. KernelExplainer) work on tree backends for free.
• New xgboost extra (xgboost>=2.0), kept out of the transparency umbrella. CI test sync installs --extra xgboost so the acceptance tests run rather than silently skip.

Follow-ups (filed)

• Add scikit-learn tree backend (TabularTreeBackend subclass) #295 — scikit-learn tree backend
• Add LightGBM tree backend (TabularTreeBackend subclass) #296 — LightGBM tree backend
• Tree/tabular backend: regression task support #297 — regression task support (P1; classification-only here)
• Add SHAP GPUTreeExplainer (GPU TreeSHAP; depends on #246) #274 — GPU TreeSHAP (depends on this)

Checklist

• CI — Required checks are green
• Breaking changes — If this PR breaks compatibility (API, configs, file formats, etc.), the title uses ! after the type or scope (e.g. feat!: or feat(transparency)!:). I understand that this change will have a direct, disruptive impact on package users. I ensured that this change is absolutely necessary and cannot be delayed.
• Contributor guide — I’ve read Pull requests and commit messages before requesting review.

Optional

• Issue (optional) — A GitHub issues is linked to this PR ("Development" section in the right sidebar).
• Docs (optional) — User or contributor docs updated where needed.
• Tests (optional) — New or updated tests cover the change.

🤖 Generated with Claude Code

---

Follow-on: registry-driven deps inference (rode along)

While fixing the .ubj → xgboost deps mapping, the extension→extra logic in deps/inference.py was a hardcoded if/elif duplicating the backend registry. Refactored so backends declare extra=/supported_hardware= on @register, and deps auto-derives the install hint via an import-free AST scan (mirrors scan_adapter_extras). Also centralised the resolved-hardware vocabulary into a ResolvedHardware StrEnum (raitap/types.py) holding all three axes (resolved name ↔ pyproject extra suffix ↔ config bucket), deleting the scattered _HARDWARE_SUFFIX dict + duplicated Hardware = Literal[...]. Separable commit (refactor(deps): …) if the maintainer prefers it split out.