fix: stop the raw install-if-different check at the end of the object

[otavio]

What

An object installed with the raw handler could be skipped as "unchanged" on the strength of a different object's version string, leaving the old binary on the device while everything around it updated.

The failure

An i.MX6ULL package writes SPL at seek 1 and u-boot.img at seek 64 of the same /dev/mtdblock1, both guarded by the same U-Boot version. Install::handle sorts objects largest first, so U-Boot proper is written first. The SPL check then ran with count: -1, which handed should_skip_install a reader over the whole device — it read straight past the SPL's own 56320 bytes into the U-Boot image written moments earlier, found the version it was looking for, and skipped the SPL.

Downgrading a device left it with the new SPL and everything else rolled back.

The fix

Bound the check to the bytes the install writes:

let region_len = if self.compressed {
    self.required_uncompressed_size
} else {
    let written = self.size.saturating_sub(skip);
    count.map_or(written, |n| n.min(written))
};

The two cases differ because skip and count measure the source, not the region it lands in: for a compressed object the source is compressed, so narrowing an uncompressed length by a compressed byte count would under-bound the check. Converting count to a byte length once, up front, also lets the write path share it, which drops the self.count.clone() that only existed to allow matching the enum twice. With both arms of the check now yielding the same type, the local AsyncReadSeek trait and its Box<dyn> go away.

flash.rs and imxkobs.rs were left alone deliberately. Flash erases and rewrites the whole MTD partition, so its check already covers exactly its own region, and a size bound would cut off the banner on NAND where nandwrite -p skips bad blocks. Imxkobs reads /dev/mtdXro because kobs-ng writes FCB/DBBT plus firmware copies at layout-defined offsets, so there is no region expressible as seek + size.

Tests

• install_if_different_stops_at_the_end_of_the_object — object at SHARED_OBJECT_AT, neighbour at SHARED_NEIGHBOUR_AT carrying the version the object asks for; the object must still be written. Fails against the previous code with "the neighbour's version answered for this object and skipped it".
• install_if_different_matches_inside_the_object — the counterpart, so a matching version inside the object's own region still skips.

The dependency bump

The second commit pins find-binary-version to 0.5.2 rather than accepting any 0.5. 0.5.0 read binaries in fixed 0x200 windows with no overlap, so a version banner crossing a boundary was never found and the binary reported no version at all — which check_if_different reads as "different", quietly turning install-if-different into "always install" for whole builds at a time. 0.5.1 fixes that; 0.5.2 additionally stops the custom-pattern search from reading the whole source into memory before matching.

Both halves are needed here: the bound alone made the downgrade work only because nothing was detectable, and the detection fix alone would still have let U-Boot proper answer for the SPL.

Verified against the real packages, driving objects::Raw::install over a simulated /dev/mtdblock1:

kirkstone (fresh flash)            SPL: written
kirkstone (re-install, must skip)  SPL: skipped
wrynose (upgrade)                  SPL: written
wrynose (re-install, must skip)    SPL: skipped
kirkstone (DOWNGRADE)              SPL: written   <- the reported bug