Everything that changed, session by session.

The pass turned on every layer one by one: 70 of 82 returned data, and the rest were legitimately empty (no aurora, markets closed on Sunday) except four.

• BGP activity is back. RIPE RIS Live stopped matching the list form of its subscription filter.
• HoneyDB points at its current endpoint (the old one now returns 404).
• Malware URLs (URLhaus) is set to “soon”: abuse.ch retired its public downloads and now requires an API key.
• Wikipedia edits is set to “soon”: Wikipedia replaced IP editors with temporary accounts, so there is nothing left to place on a map.

Also fixed: the satellite layers no longer start a new multi-MB download on every one-second tick while the first one is still running. The service worker no longer keeps a copy of every shared link. A robots.txt was added.

Thermal detections were drawn as metre-radius circles (1.8–9 km) that overlapped into orange blobs smearing over the basemap at city zoom. Now rendered as small crisp dots (~1 VIIRS pixel), legible at every zoom without disappearing.

A deploy re-applied a stale nginx.conf that dropped the local listener the Cloudflare tunnel connects to, so Cloudflare returned 502 for all dynamic content (cached static assets still served). Restored the listener and fixed it in the source config so future deploys can't drop it again.

paintPanel() rebuilt the whole 75-row DOM on every refresh (lightning every 8 s), yanking the user's scroll back to the top mid-read. Split into a structural paint + a lightweight refreshLayerRow() that only rewrites a single row's count/active state — scroll position is now preserved. The live count also flashes when a layer updates.

Four globe bugs caught + patched live: (1) points stuck at pre-zoom positions when a URL with deep zoom loaded — fixed via self-jumpTo + zoomend repaint trigger; (2) zoom slider/label desynced from actual zoom on URL boot — synced in _runOnceMapReady; (3) pulse rings + halos dominated the view at city zoom — added zoom-dependent radiusMaxPixels caps; (4) far-side bleed-through of thousands of satellites — added great-circle culling against camera center for points + rings, skipped for arcs/paths.

Critical fix. The blocking flock(LOCK_EX) added to proxy.php in [s16] caused every FPM worker to queue behind a leader fetching upstream (up to 25 s). With 5 workers + the thundering-herd-prone cables endpoint, the entire pool froze and every PHP endpoint — feedback.php, visitors.php, stats.php, ALL of them — started returning 504 because no worker was free to accept the next connection. Switched to LOCK_EX | LOCK_NB: contender either serves stale cache or 503's with Retry-After instead of waiting.

User shipped a screenshot showing a second, offset sphere of points floating above the main Earth globe. Reproduced in browser: with interleaved: true on a globe-projected map, the deck.gl overlay can finish a drag in a half-redrawn state — one hemisphere shows the pre-drag pointcloud, the other shows the post-drag one, with a visible vertical seam down the middle of the sphere. With dense layers active (satellites + Starlink = ~7 700 points), the seam looks like two separate globes overlapping.

Root cause: deck.gl's MapboxOverlay caches projection state per gesture frame; when MapLibre's transform settles at moveend, deck doesn't always know to invalidate its previous frame. Forcing map.triggerRepaint() on moveend drops the stale frame and re-projects deck.gl against MapLibre's final camera. Cheap (one repaint per gesture end) and doesn't fight the in-flight drag the way the [s18]-reverted styledata-time setProjection did.

After the [s17] interleaved-overlay fix, every mouse drag on the globe barely moved the camera — a 540 px horizontal pan produced a ~0.4° longitude change instead of the expected ~137°. The deck.gl layers (earthquakes etc.) sat motionless on screen even though the basemap should have rotated underneath them.

Root cause: the map.on('styledata', …) handler was re-asserting map.setProjection({ type: 'globe' }) on every styledata event. MapLibre 5 fires styledata frequently during camera-driven re-renders (panning, zooming, tile-load progress), and each setProjection call inside the gesture resets the projection matrix mid-drag — the cumulative pan delta gets clobbered and the camera barely moves.

Fix: removed the setProjection call from the styledata handler. The projection is set once at the constructor and re-asserted once per style.load (after setStyle() for the planet/theme swap) — that's sufficient. Verified by dragging from (740, 400) to (200, 400): lng goes from 0 to 136.8, globe rotates to show Asia/Oceania, and every earthquake ring stays glued to its real geographic position (Japan, Philippines, Indonesia, PNG, Vanuatu, New Zealand all on the Ring of Fire).

app.js:670 wrote L._error directly into the layer row's innerHTML. That field comes from e.message on a caught exception (sliced to 24 chars), which ultimately originates from upstream API responses. A misbehaving proxy returning HTML-laden errors could inject script into the panel. All panel-row interpolations now go through esc() (label, desc, count, err message).

_planetSwapInFlight was cleared in map.once('style.load') only — if MapLibre rejected the style (malformed JSON, network error on tile source), the 'error' event fired first and left the flag stuck at true, locking out every future Earth/Mars swap until a page reload. Now also reset in the error handler.

startLayer() had no idempotency guard. When a layer was both in the URL state and in defaultOn, or when setReplayMode() re-entered an already-running layer, the second call called setInterval again and overwrote the Map entry. The first interval kept firing forever with no way to stop it — slow leak measurable over hours. Added early-return when state.enabled.has(L.id).

readURLState() used parseFloat(p.get('z')) with ?? null fallback. parseFloat('abc') returns NaN, and NaN ?? null stays NaN. MapLibre 5's stricter validator throws "invalid LngLat" on NaN center/zoom and the map fails to construct. A shared link like ?z=foo&lat=bar would 502 the whole page. Now filtered via Number.isFinite().

map.on('styledata') called _boostLightLabels() which called setPaintProperty on every text-bearing layer. In MapLibre 5, setPaintProperty re-fires styledata — so each boost spawned another boost. Bounded in practice (idempotent values) but burned CPU on every theme/planet swap. Added a _boostedThisStyle flag, cleared on style.load.

After the MapLibre 5 / globe upgrade, deck.gl layers (earthquakes, ISS, asteroids, visitors, lightning) rendered as a flat horizontal stripe that stretched across the entire viewport — points stayed glued to mercator screen coordinates while the basemap had moved to a sphere. Caught from a user screenshot. Root cause: MapboxOverlay defaults to interleaved: false, which rasterizes deck.gl on its own 2-D canvas using the previous mercator projection matrix and composites it on top. deck.gl docs are explicit: globe projection requires interleaved mode so the deck layers share MapLibre's WebGL context and projection. Flipped to interleaved: true. Points now stick to the sphere correctly — verified on the Pacific Ring of Fire (India/Nepal/Philippines/Vanuatu earthquakes all land on the right continents).

Connected to the live site via the Chrome MCP extension and watched what MapLibre actually did. 29× TypeError: Failed to fetch in the console: the OpenPlanetary tile CDN is hosted at cartocdn-gusc.global.ssl.fastly.net (Fastly, not *.cartocdn.com), so the CSP's img-src and connect-src rejected every tile. Added the host to both directives → Mars surface now renders (Olympia Planum, Hellas Planitia, Acidalia, Tempe Terra all show correctly).

NASA Mars Trek tiles are served WMTS-style (/z/y/x.jpg, row-major) which doesn't align with MapLibre's XYZ scheme past zoom 0. Switched to OpenPlanetary's Mars basemap (cartocdn-gusc.global.ssl.fastly.net/...) which uses true XYZ and CORS-friendly headers. Mars planet view now renders.

Two issues: (1) the location = block didn't have an explicit try_files directive so file lookup was ambiguous, and (2) Cloudflare was caching the 403→404 response from before the whitelist was in place. Fix: added try_files /changelog.html =404; and Cache-Control: no-store + CDN-Cache-Control: no-store headers so Cloudflare bypasses the edge cache.

space.js:51 — d.altitude.toFixed(0) threw when the wheretheiss.at API returned a null field (rare but observed). The whole ISS layer stopped rendering. Fix: (d.altitude ?? 0).toFixed(0) on all 4 fields (altitude, velocity, latitude, longitude).

energy.js and infra.js — p.mw.toFixed(0) would throw if WRI shipped a null capacity. Pre-coerce to mw = p.mw || 0 at the top of the map function so the layer keeps rendering even on missing values.

Mars rovers layer (planet: 'mars') is hidden in rebuildDeckLayers when on Earth, but the URL-restore loop would still call startLayer() + fetch(). The fetch did nothing useful but populated state. Two fixes:

• URL restore now skips layers whose planet doesn't match the active planet
• fetch() bails early if state.planet !== 'mars'