P2 — Performance pass

[jenny07007]

Blocked by: S1–S8

• Lighthouse mobile ≥ 90
• LCP < 2.5s, CLS < 0.1, INP < 200ms
• Bundle analysis; R3F not shipped to routes without shaders
• All loops/canvases pause when off-screen
• Verify R3F scaling guide applied: https://r3f.docs.pmnd.rs/advanced/scaling-performance

[jenny07007]

Performance pass — results

Audited on a local production build (bun run build + bun run start), Lighthouse 13.4.0 mobile, headless Chrome.

Target	Result
LCP < 2.5 s	✅ 1.7 s under real (devtools) throttling; 95 ms observed unthrottled. See note below on the simulated 4.1 s.
CLS < 0.1	✅ 0–0.009
INP < 200 ms	✅ by lab proxy (TBT 0–10 ms); confirm with field data post-launch
Bundle analysis; R3F not shipped to routes without shaders	✅ R3F/Three.js loads only with the hero globe; mobile (<1024 px), no-WebGL, and reduced-motion get a static fallback image
All loops/canvases pause when off-screen	✅ Hero globe switches to frameloop="demand" off-screen (verified ≈1810 WebGL draw-calls/s in view → 0 off-screen → resumes); Mission illustration GSAP timelines unmount off-screen (≈520 SVG mutations/s → 0 → resumes)
R3F scaling guide applied	✅ incl. earth texture resized 4096×2048 → 2048×1024 (~45 MB → ~11 MB VRAM), no visible quality loss at desktop sizes
Lighthouse mobile ≥ 90	⚠️ 99 with real throttling / 87 with simulated throttling (the method PSI uses) — pinned there by a confirmed upstream Lighthouse bug, see below

Why the simulated score shows 87

The page's LCP is the hero subtitle (text, present in the initial HTML), and it paints in the same frame as FCP — identical trace timestamps, confirmed by logging Lantern's inputs during the audit (fcpTs == lcpTs, ~95 ms observed). Lighthouse's Lantern simulator has a known bug where it pessimistically inflates the LCP estimate using unrelated resources even when LCP is the FCP paint: our simulated FCP is 1.4 s but simulated LCP is 4.1 s for the same paint event.

This is GoogleChrome/lighthouse#16539 ("Lighthouse Punishes Good Asset Preloading", open since Jun 2025). The fix — PR #16782, "compute LCP as FCP when they occur within the same frame", opened Nov 2025 — was approved by Paul Irish on Feb 2, 2026 but is still unmerged; the latest Lighthouse release (v13.4.0, Jun 9, 2026) does not include it. PSI runs the released Lighthouse, so the PSI lab score inherits the bug.

Proof it's the bug and not the page: I applied the core diff from PR #16782 (~10 lines in core/computed/metrics/lantern-largest-contentful-paint.js) to the same local Lighthouse 13.4.0 install and re-ran the identical simulated mobile audit against the same production build. Score went 87 → 100 (simulated LCP 4.1 s → 1.4 s = FCP) with zero site changes. Also tested: removing the font preloads entirely doesn't move the simulated LCP (still 4.1 s) and regresses FCP/CLS — so there's nothing site-side left to chase.

Site-side changes made during this pass

• Off-screen pausing for hero globe + Mission illustration loops (use-in-view.ts)
• Earth texture resized 4096×2048 → 2048×1024 (~45 MB → ~11 MB VRAM, no visible quality loss)
• Open Sans font-display: optional — eliminates the late font-swap repaint that previously was the LCP under real throttling (LCP now equals FCP)

Conclusion

Every metric target passes; the only number below 90 is the simulated lab score, a measurement artifact with an approved upstream fix pending since Feb 2026. It will self-correct to ~100 when the fix ships in a Lighthouse release and PSI picks it up. Field CWV (CrUX) — what actually feeds rankings — is unaffected. Recommend closing this as complete, with a tracking note to re-check PSI after the next Lighthouse release.