The main purpose of speedrunning any category of Majora's Mask is to do what ever give give you the best PB in the long run, not what will make you get your next PB as soon as possible.
Practicing[]
It's good to make yourself as consistent on each part of the run as possible. One way to help do that is to determine which parts need a lot of practice to become more consistent. One way to learn which parts need a lot of practice might be not resetting any speedrun you're in the middle of unless you make such a big mistake that you can no longer recover the route you were using later in the run. In Majora's Mask, you can always recover the route by continuing from the last place you played the Song of Time if you don't have a faster way. You might get a better single segment PB in the long run if you do a segmented run practicing each hard part until you become very consistent at it. Also, don't reset your run just because you got 20 seconds behind unless you're quite early in the run. Speedrunning is sort of like a random walk in how far ahead you are. Suppose you take a million step random walk again and again and each time see how high a number you get to at the end and call it a record each time you get to a higher number than ever before. Then the expected number of times for you to get a record varies as the log of the number of attempts you did. The standard deviation of the probability density of landing on each number is about 1,000. If your current record is 20,000. The expected number for your next record is 20,050. However, the math shows that given that your random walk ended up at 20,000, the probability distribution of how far you were half way is normal and still has a standard deviation of 500. Although your next record is only expected to be 20,050, it's not reset worthy if you advance only 9,500 steps in the first half. It's so much more common than advancing 10,000 in the first half that the probability of advancing 9,500 in the first half then 10,500 in the second half is only square root of e times lower than advancing 10,000 in both halves. The rare time that you advance 10,500 steps in the first half, you have a pretty low chance of even advancing 10,000 let alone 10,500 steps in the second half. That's why once your record is 20,000, the expected number for your next record is only 20,050.
There might be some parts of a run that you think are RNG but maybe if you use the void warp cheat to continuously practice it, you might find a more consistent strategy and discover you can more easily make up for the RNG that happened earlier than you thought. For example, you might want to continuously practice the Twinmold fight and and the Majora fight if you're doing any% and practice the Goron race with single magic if you're doing 100%. You're probably less likely to get hit by one of the dragons when you're standing on the platform because it can't spawn under you. For the Goron race, it's more important to be more consistent at winning the race at all than to do the race as fast as you can. It's probably easier to manipulate the goron RNG if you don't loop around at the beginning of the race making you more likely to get a faster race, let alone win the race at all when you don't loop around at the beginning.
Routing[]
It's possible that the only way to get a really good PB might require you to make decisions that adapt to mistakes and RNG that happened earlier in the run. For 100%, there might be many ways you could encounter the first situation you have to adapt to and for most of those ways, there might be many ways to encounter a second situation you have to adapt to. If that's the case, it's way too long to write a route that tells you how to adapt to each situation, and it's better to become smart and figure out how to get better at adapting to unexpected situations. For example, if you accidentally wasted too many bombchus, pause the game and take the time to think what would be the best way to make up for that mistake, even if it takes a long time to do. With practice, you will get faster at figuring out the best way to make up for a mistake. In fact, the best way to speedrun, especially for 100%, probably doesn't use a route in the first place that assumes you won't make any mistakes but rather in each run, the best way to do any speedrun probably at each point in the run adapts to earlier parts of the run in such a way that you'll be less likely to lose as much time later to mistakes in the first place.
Sometimes there's a risky strat that you can do or not do where you can't possibly lose more than a tiny bit more time than the time taken to recover from the fail. If there are many situations like that in a run, you want to aim to get the average lost time to those fails low. Maybe there are so many of them that there's no hope that you will not fail any of them in a run where you do all of them. For example, in 100%, don't loop around at the beginning of the goron race because if you don't it's easier to manipulate the goron RNG and be less likely to lose the goron race. As another example, in 100%, after the goron race, you can save about 3 seconds by doing a bombchu recoil flip to the island with Tingle over it but if you do it, there's about 1 chance in 5 that you will get hit by a guay costing about 8 seconds. Therefore you should do the bombchu recoil flip. By doing it that way, you will save more time from doing the risky strats you didn't fail than you will lose from doing the risky strats you did fail.
100%[]
Since the run will probably have a slow Dampe grave, it's probably better to plan a route that assumes you will get a slow Dampe grave unless the amount of time you would lose if you used a route that assumes you will get a fast Dampe grave then you get a slow Dampe grave is very small. Instead, if you happen to be lucky and get a fast dampe grave, figure out the best way to take advantage of it after it happens. It's probably also better to use a route that gets the Postman heart piece after the bunnyhood if you're using the Japanese version and gets the dog heart piece after the Mask of Truth.