Add metric guard

[ncborcherding]

Summary

Fixes two coupled failure modes in AINet clustering: a silent collapse to a single cluster when affinity and distance use different geometries, and prototypes that drift far off the data manifold and stop corresponding to the
data.

Problem

AINet matures antibodies to maximize affinityFunc but assigns clusters with distFunc. With cosine affinity (which scores angle only and never constrainsmagnitude) paired with a Euclidean distance, antibodies random-walk to ~18x the
data scale. On WDBC this collapsed every point onto one antibody (k=1, ARI 0, antibody L2 norm ~88 vs data ~5). The grid still ranked these configs best by silhouette, so selection rewarded the degenerate runs and the interpretability
figure showed prototypes nowhere near the data.

Changes

• Metric guard (initialize): a cosine vs non-cosine mismatch between affinityFunc and distFunc warns and aligns the distance to the affinity's natural metric (cosine to cosine, otherwise Euclidean-family). Prevents the collapse.
• Consolidation M-step (consolidate, consolidationSteps, clustering only, default on): seeds an assignment by affinity, then runs Euclidean Lloyd refinement so prototypes become true data centroids. The repertoire keeps the matured antibodies and metadata, so memory and isotype state are untouched. Only result$antibodies reports the consolidated prototypes.

Validation

• WDBC prototypes now sit on the data (antibody norm ~5 vs data ~5, nearest-data distance ~2 vs data NN ~2.5).
• Gaussian affinity remains the stronger clustering choice (ARI 0.33 vs 0.14 for cosine under the same setup); the guard makes cosine safe rather than best.
• Full testthat suite passes.

Notes

• Default behavior changes (consolidate=TRUE, guard may override distFunc),
  so downstream benchmarks must be regenerated.
• Set consolidate=FALSE to recover the raw affinity-assignment behavior.