[casoon]

Package 2, part 2: performance. This builds on the collection contract from #799.

#770 Measurement windows

• LCP: the last candidate before the read, including candidates painted after load. The old cut-off at load + 100 ms kept the small initial text as LCP.
• TTI: a lab proxy of the Lighthouse definition. It is the end of the last long task before the first 5 s quiet window after FCP, and at least domInteractive. It stays marked estimated_lab because network quiet is not observed.
• TBT: the blocking time of long tasks over FCP→TTI. Each task is clipped to that window, and the part beyond 50 ms counts. It is computed in Rust from the raw long-task list. Before, tasks were summed up to load + 100 ms.
• Collection end: collection stops after 1 s without a new long task (Lighthouse cpuQuietThresholdMs), or after 8 s at most.
• Speed Index: stays a diagnostic estimate without score weight.
• Removed tbt_zero_heavy_page: the warning claimed TBT = 0 with LCP > 3 s is impossible. That is false: a static page on a slow network has no long tasks. The locale keys are removed too.

#769 Measurement evidence

• Records:
  • Each core metric records its own state: performance.lcp, performance.fcp, performance.cls, performance.tbt. The states are measured, failed (observer_not_installed, metric_not_captured, fcp_not_captured, invalid_metric_value) and unsupported (entry_type_not_supported).
  • The observers, CDP metrics and navigation timing record their state as well.
• Sub-analyses decode strictly: content weight, render blocking, third party, critical chain, minification and animations. A failed read is no 0-byte page and no clean list. Content weight feeds the score, so losing it makes performance partial.
• Coverage: with no profiler answer, JS coverage reports used_pct: null and js_coverage_unavailable instead of 100 % used.
• Third-party isolation: an origin whose isolation pass failed makes the collection partial or failed instead of silently going missing.
• Throttled profiles:
  • A profile whose CPU throttling or cache disabling failed is no longer stored under that profile's name.
  • A score requires all four core metrics; otherwise it is score: null, reported as "incomplete" in the PDF.
  • Only a complete LhMobile profile becomes canonical. The old min(50) score for a missing LCP is gone.

Tests

• Unit tests:
  • TTI quiet window;
  • TBT FCP→TTI clipping;
  • late task after load;
  • slow static page without a warning;
  • per-metric states, including no observer, unsupported and NaN;
  • coverage and isolation states.
• New browser tests: lcp_includes_a_hero_rendered_after_load, tbt_counts_a_long_task_after_load and slow_static_page_measures_zero_blocking_time (Slow 3G with a render-blocking stylesheet and no script).
• The existing foundation_missing_observers_are_not_perfect_zero_metrics test now also checks the reasons.

Validation (local, all green)

• fmt passes, and clippy -D warnings is clean with all features and with no default features.
• cargo test --all-features passes.
• The ignored suites CI runs all pass:
  • integration_test, 45/45;
  • detection_corpus_test, schema_rules_detection_corpus_test, vulnerable_libs_detection_corpus_test;
  • easy_language_detection_test, help_mechanisms_detection_test, score_calibration_test.
• reference_sites_test shows no deviation.

Browser runs take a little longer because of the 1 s CPU-quiet wait. The detection corpus went from about 400 s to 463 s.