fix(fires): stop discarding 96.8% of FIRMS detections

[xer0decibel]

The CSV parser capped output at 2,000 points by keeping every Nth row. The global 24h feed carries ~61,500 detections, so it kept every 31st — and kept them by position in the file, which is arbitrary, giving any individual fire roughly a three-in-a-hundred chance of appearing. Fires that plainly existed were simply absent from the map, and it read as a data problem rather than a parser one.

Nothing is sampled now. Rows are sent column-oriented instead of one object per detection, which is what makes that affordable: the same 61,500 are 7.7MB as objects against 2.1MB this way, ~0.6MB gzipped. lib/fires expands them back on the client, and its tests pin the column order — a mismatch would not throw, it would read longitude as latitude and scatter every fire into the sea.

Also wrapped in cachedSource. The fetch pulls 4.8MB and parses 61,500 rows, which took ~19s cold and repeated on every miss; now only the first caller pays and a failed refresh serves the last good index rather than emptying the map.

[vercel]

@xer0decibel is attempting to deploy a commit to the Developing Osiris' projects Team on Vercel.

A member of the Team first needs to authorize it.