The actual damaged 64K age blocks would likely be unrecoverable after the start of the damage unless the chacha20-poly1305 ended being self synchronizing as it was used (as opposed to the CFB that OpenPGP specifies). The question I can't answer is if the undamaged 64K age blocks would then be recoverable. There might be just a counter, but you could instead (also) make a particular block dependent on the previous one(s).

Do you understand that chacha20 is CTR? A bit flip only affects that single bit, does not propagate to any other bits?

From that can I get that an age recovery utility would need to detect missing data and would then need to insert dummy blocks (or the equivalent)? I guess there would have to be an minor element of brute force involved as there would be no easy way to distinguish bad blocks from the blocks after the missing chunk.