nx_gpt_relocate.nx

nx_gpt_relocate.nx -- FIX THE WRITE, NOT THE IMAGE (debt 1785972535). An image is a FIXED-SIZE disk. A USB stick is usually BIGGER. So a raw byte-for-byte write leaves the BACKUP GPT sitting in the middle of the device instead of at its final LBA, and the primary header still describes the image's geometry. Every gate we had passed because every gate judged the IMAGE -- but the artifact that ships is the DEVICE. Measured on a real stick: 7,987,200 device sectors vs 135,269 image sectors; the disk came up GPT with the ESP visible and the data partition missing entirely. This organ makes a written device's GPT describe THAT DEVICE: * primary header : AlternateLBA = last LBA, LastUsableLBA = last - 33, header CRC32 recomputed * backup entry array: copied to (last - 32) * backup header : written at the last LBA, MyLBA/AlternateLBA/PartitionEntryLBA corrected * the STALE backup : zeroed where it used to sit -- a leftover secondary GPT mid-device is a landmine, because a future repair tool can "restore" the old geometry from it and silently shrink the disk back. It REFUSES rather than corrupting: no "EFI PART" signature, a device smaller than the image, or a partition that would extend past the new last-usable LBA all stop before a single byte is written. HOST-INDEPENDENT BY CONSTRUCTION: it seeks and writes a block device or a file through plain syscalls, so the same organ serves Linux, Nishi OS itself, and any host that can run our ELF; hosts that cannot (Windows today) apply the identical byte ranges, and `plan` prints exactly those ranges so the two can be compared instead of trusted. One brain, many hands. usage: nx_gpt_relocate selftest nx_gpt_relocate plan <path> <device_sectors> (prints the byte ranges, writes nothing) nx_gpt_relocate apply <path> <device_sectors> (writes, then RE-READS and verifies) license_tier: ORIGINAL expect_exit: 0