Scope wall-clock selection budgets per strategy instead of per run

[lullu57]

Summary

wall_clock_limit_s starts once for the whole run, before compile-signal collection, and is then shared by all strategy brokers. Later strategies can be penalized because earlier strategies or pre-selection work consumed the clock.

Evidence

• src/kernel_tuner/experiments/orchestrator.py:318
• src/kernel_tuner/experiments/orchestrator.py:522
• src/kernel_tuner/experiments/orchestrator.py:561
• src/kernel_tuner/experiments/orchestrator.py:603
• docs/04_experiment_protocol.md:45

Why This Matters

The protocol defines this as a strategy selection budget. Sharing one wall-clock deadline across the entire run makes cross-strategy comparisons unfair.

Expected Fix

• Scope wall-clock accounting to each strategy decision, not the entire run.
• Exclude pre-selection work like compile-signal collection from per-strategy selection deadlines unless explicitly documented otherwise.
• Preserve explicit skipped_budget records when a strategy genuinely runs out of time.

Acceptance Criteria

• Each selector/baseline starts with the same wall-clock selection budget.
• Pre-selection phases do not silently consume strategy budget.
• Tests prove one strategy cannot starve later strategies of time.