perf(flights): skip dead-reckoning contacts that stay behind the horizon

The fleet tick dead-reckoned, floored and distance-tested every contact on
every ~80 ms tick, then hid the ones behind the horizon. The worldwide feed
carries ~13k contacts; zoomed in, well under 1% are on screen, so nearly all
of that work produced a hidden billboard.

A contact that is already hidden and still behind the horizon at its last
position now re-reckons only on its turn in a staggered one-in-12-ticks
rotation (about once a second, HIDDEN_CONTACT_DR_TICKS). Anything that
position puts in view, and any contact with a fleet model, takes the full
pass immediately, so on-screen behaviour is unchanged. Hidden-inclusive
proximity queries (getNearby includeHidden, getAllPositions) read positions
at most about a second old, well inside the 30 s poll's uncertainty.

Measured in Chrome on a Radeon 780M, ~12.9k contacts, 56 on screen over
Colombia at 4 km, same view before and after:
  fleet tick   7.66 -> 2.03 ms/frame, longest tick 50.9 -> ~16 ms
  frame        20.6 -> 14.0 ms, 42.9 -> 57.8 FPS
Teleporting the camera to Madrid showed 975 contacts within 100 ms.

Co-Authored-By: Claude Opus 5.5 <[email protected]>

docs/CURRENT-STATE.md

@@ -3469,7 +3469,7 @@ silently demoting every later lookup for the session.
 
 ### Motion & Symbology Correctness (June 10, 2026)
 
-- **Flights (commercial + military)** render one poll interval behind real time (30s/15s) and interpolate between two known feed-stamped fixes (OpenSky `time_position`; adsb.lol `receipt − seen_pos`). The display latency is an intentional product decision — do not "fix" it away. The whole fleet dead-reckons at ~12Hz (1m² write gating); aircraft get a 3-poll grace period (faded icon) before removal. When a position epoch pauses, both layers coast for at least 60 seconds of contact grace with an absolute five-minute ceiling. Source backoff marks each contact and the cockpit `STALE`; the cockpit then holds the exact layer position instead of continuing inertial flight. Repeated-position kinematic changes create a forward-only synthetic fix rather than mutating history, and grounded history is lifted only when no owned 3D model already controls its datum. A nominally successful worldwide OpenSky response whose own snapshot epoch is more than two minutes old prefers the existing 250-nm viewport-scoped adsb.lol fallback and labels the source/coverage accordingly when that upstream is available. If the fallback is also unavailable, the layer's freshness/error fields use the source epoch—never the cache receipt time—so the UI reports an old snapshot rather than “just now.”
+- **Flights (commercial + military)** render one poll interval behind real time (30s/15s) and interpolate between two known feed-stamped fixes (OpenSky `time_position`; adsb.lol `receipt − seen_pos`). The display latency is an intentional product decision — do not "fix" it away. The fleet dead-reckons at ~12Hz (1m² write gating); a contact already hidden behind the horizon whose last position still is re-reckons on a staggered one-in-12-ticks rotation (about once a second) instead, so the worldwide feed's ~13k contacts no longer cost a full pass when well under 1% are on screen. Anything that position puts in view, and any contact with a fleet model, takes the full pass immediately, and hidden-inclusive proximity queries read positions at most about a second old; aircraft get a 3-poll grace period (faded icon) before removal. When a position epoch pauses, both layers coast for at least 60 seconds of contact grace with an absolute five-minute ceiling. Source backoff marks each contact and the cockpit `STALE`; the cockpit then holds the exact layer position instead of continuing inertial flight. Repeated-position kinematic changes create a forward-only synthetic fix rather than mutating history, and grounded history is lifted only when no owned 3D model already controls its datum. A nominally successful worldwide OpenSky response whose own snapshot epoch is more than two minutes old prefers the existing 250-nm viewport-scoped adsb.lol fallback and labels the source/coverage accordingly when that upstream is available. If the fallback is also unavailable, the layer's freshness/error fields use the source epoch—never the cache receipt time—so the UI reports an old snapshot rather than “just now.”
 - **World-space headings at every angle** (`src/data/iconOrientation.js`): aircraft/vessel icon rotation uses the camera right/up basis per tick (alignedAxis always ZERO), which is exact at screen center and for orthographic/nadir views and remains stable through >180° tracked orbits. Perspective rays vary across the viewport, so off-center contacts at oblique pitch can diverge from an exact finite-difference window projection; a regression test pins that known regime, and field evidence decides whether to adopt exact projection with the basis method as fallback. Fleet rotations refresh on camera-pose change; tracked entities per frame. Billboards are horizon-culled via a shared EllipsoidalOccluder.
 - **Military/OpenSky reconciliation** (`src/data/militaryRegistry.js`): known-military ICAOs render amber in the flights layer (60s self-poll of the cached mil endpoint when the military layer is off) and are suppressed there while the military layer renders them.
 - **Satellites**: 838-sat core catalog (stations/visual/GPS/GLONASS/Galileo/GEO), tracking lands ~726km out via `viewFrom` (tracked entity owns a point graphic so the tracking camera engages), rings realign via primitive modelMatrix (no per-second rebuild flicker), optional `setParams({catalog:'dense'})` Starlink mode.