The only time where I have seen this done is where there were several different name-spaces in play, and they did have a visibility-hierarchy relative to one another (such as local vs. global symbols), and the entirety of a name-space was created and then thrown away many times as the program ran. You unquestionably could human-notice the time that it took to run through all the hashes on a hash-miss, but it was a price to be paid.
Absent such considerations, though, I generally approach this problem by simply creating one hash into which every symbol encountered is added, but never removed. The symbol’s hash-entry points to a record/object in which the various meanings-of-the-moment are maintained in a list. (A symbol might appear in the hash but have no current meanings at all, and in the application that I am thinking about, the same names tended to be re-used many times. The price of insertion was paid but once, and the price of removal from the hash, not at all.)