feat: ShapModelPlot.clustermap (explanation-similarity) + shap_to_feat_imp + sample= SHAP plots

[breimanntools]

Part of #305.

Problem

The richest explanation figure in the original project — clustering samples by
explanation similarity (correlation of their per-sample SHAP-value vectors,
so proteins group by why the model scores them, not just by features) — is
~130 lines of sns.clustermap + scipy.cluster.hierarchy surgery
(plot_shap_cluster_map / plot_shap_dendrogram, scripts/plot_cpp_shap.py).
It has no library API. Separately, per-sample SHAP plots in the γ-secretase
notebook (cells 30/32) couple the plot call to add_feat_impact's naming by
hand:

seq_kws = sf.get_seq_kws(df_seq=_df_gsec, df_parts=df_parts_dom, sample=acc)
fig, ax = aa.CPPPlot().feature_map(df_feat=df_feat, shap_plot=True,
                                   col_imp=f"feat_impact_{name}", **seq_kws)

The col_imp=f"feat_impact_{name}" string-templating + get_seq_kws plumbing is
repeated per sample.

Goal

Ship a SHAP-correlation clustermap and let the existing SHAP plot paths resolve a
sample by name, removing both the 130-line clustermap boilerplate and the
per-sample string plumbing. All SHAP pieces stay in the pro extra.

Requirements

• ShapModelPlot.clustermap(...) — correlation clustermap of per-sample SHAP
  vectors with row/col class-color sidebars (port plot_shap_cluster_map).
• shap_to_feat_imp helper: SHAP vector → normalized signed feature impact
  (shap/Σ|shap|*100) + absolute importance (port ml_shap.py); reuse any
  existing ShapModel logic to avoid divergence.
• Let feature_map / ranking / profile accept sample="APP" to resolve
  the feat_impact_* column + seq_kws internally (removes the
  get_seq_kws + col_imp plumbing from cells 30/32).
• numpydoc + Examples include; pro-gated import.

KPIs / Acceptance criteria

• clustermap reproduces the project figure's clustering on the canonical
  SHAP matrix (linkage + cluster assignment match).
• feature_map(..., sample="APP") equals today's explicit
  col_imp=f"feat_impact_APP" + seq_kws call (regression).
• shap_to_feat_imp matches the manual normalization within float tolerance.
• ≥1 unit test per new symbol; pro import-gating verified (core import
  still succeeds without SHAP).

Scope / non-goals

• pro only (SHAP is a pro dep). The custom trimmed/recolored sub-tree
  rendering is optional polish, not required for the KPI.
• Excludes the FPDF multi-panel report (out of scope per epic epic: use-case convenience methods — collapse γ-secretase notebook boilerplate #305).

Dependencies

• Part of epic epic: use-case convenience methods — collapse γ-secretase notebook boilerplate #305 · relates Visualization layer (multi-scale explanation) #56 (visualization layer).

Standards checklist

• pro/core gating · CONFIRM-FIRST (__init__.py/__all__) · plotting.md
  compliance · numpydoc · tests · no-print

[breimanntools]

Delivered: the explanation-similarity clustermap shipped as AAPredPlot.clustermap (#332); the per-sample SHAP sample= plotting on CPPPlot.ranking/profile/feature_map and shap_to_feat_imp shipped via #323. Note: shap_to_feat_imp currently lives in aaanalysis.explainable_ai_pro (importable there); its top-level re-export was intentionally deferred.