The problem was introduced in 5.5.20 by Bug 13116225.  It tried to
protect against downgrading from a version 5.6.4 database that was
created with a page size other than 16k.  Version 5.6.4 supports
page sizes 4k and 8k and stamps that page size into the header page
of each tablespace file.  Version 5.5.20 attempts to read that page
size in the file header.

But it turns out that only the first system tablespace file has a
reliable flags field in the header. So only ibdata1 can be or needs
to be tested for another page size. Extra system tablespace files
like ibdata2, ibdata3, etc do not and should not be tested since the
flags field is unreliable.