perf(frontend): cache static deck.gl layers, code-split heavy views, vendor chunking

[d3mocide]

Summary

Two-front performance pass on the frontend: stop the 3D globe's render loop from rebuilding the world every frame, and stop the initial page load from shipping every map/chart engine up front. A third commit fixes the satellite "catch-up surge" on globe load.

Rendering (globe frame time)

The useAnimationLoop rAF loop was calling composeAllLayers() at 60 fps, constructing ~40–50 new deck.gl Layer instances per frame — including static groups (cables, towers, airspace, NWS, GDELT, clusters, …) whose data changes every 30 s–15 min. Several builders also allocated new data arrays per frame, forcing deck.gl to regenerate and re-upload GPU attribute buffers continuously.

• New LayerCache (src/layers/layerCache.ts): every static layer group in composeAllLayers() is memoized on a deps array (React-hooks-style Object.is contract). Unchanged groups return the identical Layer instances, so deck.gl skips diffing and attribute work for them outright. Only entity/satellite/trail layers — whose positions genuinely change each frame — still rebuild per frame. Each overlay owns its own cache (Layer objects hold per-deck state).
• Pulse animations quantized to 10 Hz (aurora, jamming, FIRMS, dark vessels, holding patterns, kiwi node): visually identical shimmer, 6× fewer attribute recomputes.
• Terminator memoized per minute — it was recomputing a 360-point polygon and allocating a fresh GeoJSON (= full GPU re-upload) 60×/sec.
• StarField uses globalAlpha instead of formatting ~320 rgba() strings per frame; interpolation hot path no longer recomputes trig constants per entity.

Loading (initial latency)

• Code-split the five heavy views (TacticalMap, OrbitalMap, IntelGlobe, DashboardView, RadioTerminal) via a lazyView() helper with built-in Suspense — the view-switch JSX is untouched.
• Vendor chunking in vite.config.ts (function-form manualChunks): deck-gl (1.03 MB), maplibre (1.07 MB), mapbox (1.72 MB), echarts (1.14 MB), react-vendor. Vite's virtual preload helper is isolated into its own chunk — otherwise Rollup parks it inside a vendor chunk and the entry transitively preloads that vendor (the object form of manualChunks has the same failure mode, which is why the function form is used).
• Map CSS split: each adapter imports only its own library's stylesheet, so the unused engine's CSS never downloads (previously both were always bundled).
• Fonts: consolidated into one index.html link resolving in parallel via preconnect; removed the render-blocking @import from the bundled CSS.

Result: the entry now modulepreloads only preload-helper (~1 kB) + react-vendor (60 kB gzip) — down from an initial graph that eagerly included all of deck.gl/maplibre/mapbox (~5 MB+). Vendor chunks are byte-stable across releases, so returning clients hit the HTTP cache.

Satellite catch-up surge fix

Satellites loaded, drifted slowly along their orbits, then periodically all surged across the globe at once. Root causes and fixes:

• Epoch-anchored dead reckoning. Satellite messages carry time (the SGP4 propagation epoch), but the client anchored DR to receive time — sweep chunking + Kafka + WS delivery make positions 1–10 s old on arrival, so satellites rendered persistently behind reality until the next 15 s sweep "corrected" them. DRState now splits serverTime (position epoch — projection anchor) from blendTime (receive time — blend-progress anchor); new satellites appear at their true, epoch-projected current position immediately. Guarded against clock skew / wrong units (drAnchorTime()).
• Teleport guard in interpolatePVB: visual→target gaps beyond max(2°, speed × interval × 3) snap in one frame instead of racing across the map; the longitude delta is deliberately unwrapped so antimeridian crossings snap to the far side rather than smoothing the long way around the globe.
• Frame-stall reset: after a >1 s rAF gap (hidden tab, GC, shader compilation — the original trigger, amplified by the per-frame layer rebuild this PR also fixes), visual states are cleared so everything re-seeds at its target in one clean snap instead of a global catch-up surge.
• Satellites also gain out-of-order/duplicate sweep rejection and a correct 15 s expectedInterval seed.

Verification

• pnpm run lint — clean (0 warnings)
• pnpm run typecheck — clean
• pnpm run test — 278/278 passed (6 new tests covering epoch anchoring and the teleport guard)
• pnpm run build — chunk graph inspected in dist/: vendors are on-demand, entry preloads nothing heavy, mapbox vs maplibre load exclusively

Follow-ups (out of scope)

• public/world-countries.json is 14 MB of raw GeoJSON fetched at runtime — simplifying geometry (mapshaper/TopoJSON) would cut multi-second map start on slow links.
• Entity interpolation is still O(n)/frame on the main thread; moving it into the existing TAK worker is the next rendering win if entity counts grow past ~5k.

Task logs: agent_docs/tasks/2026-06-12-frontend-performance-optimization.md, agent_docs/tasks/2026-06-12-satellite-position-jump-fix.md

https://claude.ai/code/session_01BksbmT1NKcWRSiPA9FUh6B