Corrected republish of meteolibre-dev/globalsatmetarv2. Same shards, schema, patch grid and satellite/radar/elevation bytes; only the rasterized METAR/SYNOP channel was fixed. The v2 rasterizer used a closed 1-hour window [t-1h, t], so every observation stamped at exactly HH:00:00 (all SYNOP/FM-12 reports, and any station with reporting gaps > 1h) was written into two consecutive hourly frames. ~20-40% of consecutive obs pairs in v2 were exact copies of the same report, which trains a copy-prior ("next frame = last frame") into sequence models. In v3, a frame t+1 whose tmpc, dwpc and mslp are all bit-exact equal to frame t (at least one finite) is treated as the same report leaked across…