Make profile-shape and standalone profiling target selection defensible

[lullu57]

Summary

Profile-shape sampling and standalone profiling target selection are deterministic but not necessarily representative or valid for evidence. per_workload_class_top1 selects lexicographically smallest shape_id, and the standalone profiling path profiles the first generated candidate rather than a known compile-success/correctness-passing candidate.

Evidence

• src/kernel_tuner/experiments/orchestrator.py:121
• src/kernel_tuner/experiments/orchestrator.py:130
• src/kernel_tuner/experiments/orchestrator.py:136
• src/kernel_tuner/profiling/adapter.py:252
• src/kernel_tuner/profiling/adapter.py:257
• docs/specs/profiling_adapter.md:63

Why This Matters

Tier 1 profiling is supposed to be matched-budget evidence on a calibration subset. ID-order shape selection and arbitrary standalone target selection can bias or weaken what the counters actually mean.

Expected Fix

• Make profile-shape selection explicitly representative or at least clearly documented as deterministic-only.
• Ensure standalone profiling targets only compile-success/correctness-passing candidates.
• Add tests around profile-shape and standalone-target selection.

Acceptance Criteria

• Profile-shape selection policy is explicit and defensible.
• Standalone profiling never targets an arbitrary invalid candidate.
• Tests cover both behaviors.