Sorry Irog, I skipped your question.
The trick is to hash a normalised copy of the source line rather than the raw one. Before hashing, replace every custom tag with a positional placeholder, so an icon tag and a variable tag both become tokens like {1} and {2} in the order they appear. Then keep two hashes beside the translation, one of the normalised prose and one of the tag sequence.
That splits the two kinds of change that actually matter. If the prose hash moves, the translator has real work. If only the tag hash moves, a variable or an icon was renamed or reordered, which is usually a mechanical fix and often not the translator's job at all.
It also stops a pure retag from flagging a whole file as stale, which is what made us stop hashing raw lines.