[Profile #478] Expand execution-substrate profiling from sampled operations to the full supported estimator surface

Parent: #478

The architecture audit inventories 491 estimator operations but has only 32 dynamic profile rows and many low/medium-confidence substrate labels. Static ownership labels are useful hypotheses, not enough evidence for optimization decisions.

Build automated profiling for the supported Flow-equivalent surface, capturing Python self-time, native/Cython/BLAS time, call crossings, allocations, cache/memory proxies where available, and operation phase. Prioritize all canonical operations, then all implemented estimators.

Acceptance: 100% dynamic profile coverage for canonical operations; confidence upgraded from low/medium where measured; profile artifacts keyed to sklearn version and hardware; optimization roadmap uses measured substrate when available and never labels already-equivalent as a performance conclusion.

Start by reading parent issue #478 and the architecture-audit inventory of estimator operations. Define the profiling run around canonical operations first, then implemented estimators, and record the requested timing, crossing, allocation, cache/memory, and phase data. Done means full canonical coverage, version- and hardware-keyed artifacts, upgraded measured confidence, and no performance conclusion from already-equivalent labels alone.