The 1 bp isn't liftOver disagreeing with itself, it's that the same numbers went into two different coordinate conventions. UCSC's tool reads input as BED (0-based, half-open); a GRanges is 1-based inclusive. Convert UCSC's output out of BED and its start becomes 145626812, which is exactly what rtracklayer gave you - what's left at the other end is the half-open end plus the fact that the interval you handed rtracklayer began one base earlier than the one UCSC lifted.

On the fragments, rtracklayer is showing you the truth and UCSC is hiding it. The chain really does break that region into blocks; the web tool collapses to one interval when enough of the region maps (minMatch 0.95 by default). Taking min start and max end silently spans whatever sits between the blocks, so it's only safe if they're contiguous and in order. Run reduce() on what comes back - if it collapses to a single range you're fine and it'll agree with UCSC, and if it doesn't, the region doesn't lift as a unit and UCSC's tidy answer is the misleading one.