Test budgets count CPU work, not wall time

How long a test suite takes on the clock depends on whatever else the machine is doing, so a wall-clock budget fails on a busy host even when nothing got slower. The work the suite itself does (CPU time, instructions retired) does not change with load, so budgets are counted in that work.

The user, 2026-10-15: “yes count in cpu work”.

• Test and suite budgets are counted in CPU work that host load does not affect (instructions retired or CPU time, from the sandbox’s instruments), not wall time.
• Suites may then run concurrently.
• UI-CPU extends it to the browser speed gates; NO-LOAD-BLOCK makes it the rule for every judged measurement.

Done so far: the suite budget is a CPU budget. engine/src/shared/run-cpu.ts reads a run’s CPU time, and engine/src/shared/budget-verdict.ts judges it (engine/test/prevention/suite-cpu-budget.test.ts drives that door).

Still missing (TEST-10X): the counter is process CPU time, not instructions retired, and the verdict still defers a saturated host as provisional, so the judgement is not yet fully load-independent; the measurement does not yet come from the sandbox’s instruments; suites do not yet run concurrently.