fix(metrics): correct CPU and queue alerts with bounded compile statistics

[ttbombadil]

Summary

• Correct process CPU microseconds/milliseconds conversion (95% was reported as 0.095%); retain the last valid sample across equal/backward timestamps.
• Feed the existing compile queue alert from sandbox-start, REST worker and fallback gatekeeper queues. Preserve the documented legacy sandbox-start aliases and public response shape.
• Replace the unbounded duration history and repeated reduction with constant-space lifetime total/count, preserving the existing completed-result population.

Reproduction and scope

Base: freshly fetched origin/main 4f903881267f55e32b7c94b8c13f0076bdf6a79c, after green post-merge CI for #187.

Seven regressions reproduced CPU scaling/nonfinite sampling, missed REST queue alerts, and 2,000 retained durations after 2,000 compiles. Three additional tests preserve native queue aliases, idle behavior and average sample population. Focused suites: 42 passed. Independent review: no findings.

CPU remains aggregate process CPU (100% equals one fully used core); no capacity tuning, host utilization reinterpretation or wall-clock architecture change.

Verification

• Full unit/coverage: 2,790 passed, 1 skipped; Sonar quality gate passed.
• Types, docs, non-fixing ESLint, complete build.
• Real toolchain integration: 14 passed, 1 skipped; Docker integration: 27 passed.
• Full Chromium E2E on an isolated local port: unchanged assertions; only the temporary configuration and hardcoded iframe URL adapted from 3000 to 3307 because an unrelated project owns 3000.
• Normal commit and pre-push gates; no bypass.

No Provider calls, SSOT/ADR or Tutor/Planner/Mastery/Prompt changes; protected/untracked user files remain untouched.