[Perf #478] KernelSVC predict: eliminate remaining small-shape inference loss and preserve Digits win

Parent: #478

Canonical v2 prediction is dataset-sensitive: Iris Flow predict 0.118 ms vs sklearn 0.150 ms is already faster, while current end-to-end Iris still loses because fit dominates; Digits Flow predict 17.96 ms vs sklearn 25.86 ms is also faster. The older architecture map still describes SVC.predict as a loser, so inference evidence is stale/inconsistent across artifacts.

Audit the prediction path separately from fit. Profile support-vector traversal, RBF distance calculation, vectorization, memory layout, batching and multiclass voting. Establish whether any size regime still loses and remove stale claims when it does not.

Acceptance: prediction-only scaled curves across support-vector counts/features/batch sizes; artifact generator must use current prediction evidence; no fit timing mixed into predict classification; target >=1.05x across supported inference regimes or a specific tracked loss with cause.

Start by profiling the KernelSVC prediction path separately from fit, covering support-vector traversal, RBF distance calculation, vectorization, memory layout, batching, and multiclass voting. Generate prediction-only scaled curves across support-vector counts, feature counts, and batch sizes, then update the artifact generator and stale claims based on the evidence; done means supported regimes reach >=1.05x or document a tracked loss and its cause.