Agreed. The confusing part here as I see it is where validation layer A (correctly) asserts the data is valid UTF-8 and safe then assumes the database persists what it passes to it, since no error is reported.

Then, subsystem B trusts reading the database field (since it passed validation layer A).

Obviously more validation layers can be added, but at this point validation layer C called by subsystem B needs to know what the initial input from layer A is in order to differentiate it from the db value which was manipulated - a rather tricky thing to do sometimes. (I guess you could add a hash to the db to check the db is storing your strings, but really.. come on)

Upgrading to utf8mb4 is probably safer than hoping enough validation layers thrown at it solves the problem.