The problem was in the validation of the input data for blob types.
When assigned binary data, the character blob types were only checking if 
the length of these data is a multiple of the minimum char length for the 
destination charset. 
And since e.g. UTF-8's minimum character length is 1 (becuase it's 
variable length) even byte sequences that are invalid utf-8 strings (e.g. 
wrong leading byte etc) were copied verbatim into utf-8 columns when
coming from binary strings or fields.
Storing invalid data into string columns was having all kinds of ill effects 
on code that assumed that the encoding data are valid to begin with.

Fixed by additionally checking the incoming binary string for validity when 
assigning it to a non-binary string column.
Made sure the conversions to charsets with no known "invalid" ranges 
are not covered by the extra check.
Removed trailing spaces.

Test case added.