Stabilize Lighthouse budgets with confirmation samples

[zhaowl94]

What changed

• Keep all Lighthouse category budgets at 0.95.
• Run a single audit when the first sample passes.
• When the first sample misses any budget, collect two confirmation samples and evaluate the three-run median.
• Preserve every raw HTML/JSON report and cap each audit at 90 seconds.
• Add focused unit coverage for median calculation, thresholds, retry behavior, and persistent regressions.

Why

The master run 30602101413 reported root-page performance 0.94 because Lighthouse measured 284.5 ms total blocking time on a page with no application JavaScript. All other categories/pages and prior local/PR/master runs scored 1.00, so this was runner CPU variance rather than a product regression. Confirmation sampling filters a one-off outlier without weakening the quality budget.

Impact

Normal passing CI remains fast with one sample per page. A first miss receives two independent confirmation samples; genuine regressions still fail when the median remains below 0.95.

Validation

• Windows: npm run check, 15/15 unit tests, build, 163/163 browser tests, Lighthouse 5/5 pages at 1.00, dependency audit (0 vulnerabilities), and 5/5 external links.
• Isolated Ubuntu: 15/15 unit tests, 142/142 browser tests across Chromium/Firefox/WebKit/mobile/visual projects, and Lighthouse 5/5 pages at 1.00.
• Formatting, linting, and changed-file scope checks passed.