Hi Adrian,
I came across a similar problem last year. The trouble was with the merge function of osmosis. It takes the bounding box of the output file, from the first input file, regardless of the bounding boxes of the other input files. But the data from all the input files is in the merged file. I contacted the maintainer of osmosis about this but received no reply. It may be that it is the intended behaviour, rather than a bug.
Good to know, I'll try to follow up on this.
I worked around this problem, but it is not at all convenient. You create the output file in uncompressed .osm format. This can result in a large file. [...]
Without having tried it. If - as it seems - osmosis takes the first map for the bounding box, one could create a dummy empty map with a spanning, large enough bounding box as the first one and merge it with the other ones.
I have a vague recollection that it is also possible to get the splitter to ignore the bounds in the input file.
Any more thoughts on this coming back in the meantime? Dominik