The simple answer to your question is that you type the hex-codes directly into the source records. Type into a part of the source record that is ignored by the compiler, such as into the sequence number field of a RPG statement.
The SEU display must be positioned so that the character is displayed on the screen. If you put the character into position 5 of the source statement, and SEU displays the statement starting from position 6, your workstation display won't see the hex character and won't know what color attributes to show. For that case, you might put the hex-code into position 8 of a comment line -- once the compiler knows the line is a comment, the rest of the line will be ignored.
However, the more complex answer is that it shouldn't be done unless you can guarantee that the source won't be manipulated on a different system. I.e., you do not want non-text characters to be sent across a network to a different system that may have a different system CCSID setting or to a PC that will have a different encoding scheme. You will have no good way to predict what a particular network transfer protocol will do with embedded non-text hex values, mostly because the issue won't be expected. The source may easily be corrupted in ways that can be hard to find. Migration to a new system is one example of when source might be transferred. Save/restore is usually fine, but DDM is sometimes used during transitions to phase things in.
CCSID conversions, EBCDIC<>ASCII conversions, unpredictable conversions -- I've seen where network changes brought remote access to source, and the source had to be gone through to eliminate characters that are not supported. Not fun.
Nowadays, WDSC (and RDi) are the common ways to access source remotely. But I first ran into the problem when simply sending source members to a second AS/400, back in either V2R3 or one of the V3 releases.
IMO, best would be to use delimited comment blocks. Surround critical code in some standard way such as lines of all asterisks or dashes or whatever. The time to find out that a problem exists is _NOT_ when you're in the middle of a time-critical procedure.