In each tower, the top section can be made to "glitch" (https://kb.speeddemosarchive.com/Castelian_(NES)/Game_Mechanics#Glitched_tower). While the mechanic has been known, it was only recently seen implemented in RTA runs. Presumably first by YuaRyuc, https://www.twitch.tv/videos/1902842971?t=0h7m50s.
The uploaded movie saves almost half a second in tower 5 thanks to the new finding. When continuing to TAS the remaining towers, the uploaded movie file is at one point 33 frames ahead of the current TAS (due to favorable "random" lag). However, it falls behind during the transition to the bonus level of tower 7 ("random" lag) and never recovers, clocking in 3 frames after the current TAS.
It seems like it should be possible to (e.g.) play around with the screen transitions of the earlier towers to find a combination not resulting in the extremely laggy screen transition to the 7th bonus level. However, that is nothing I plan to do.