A program for calculating the strength of the guard on random maps in Heroes of Might and Magic III, called H3CalcGuards.

Regarding the bug with the increasing number of neutral creatures, I confirm it. It turns out that the game stores the number of neutral creatures with an accuracy of up to 16th (i.e., an integer + x/16), which is why even small groups of creatures accumulate from week to week. I hadn't paid attention to this before. I will fix it and update it in a couple of days.

3 and 4. That's where I got a little confused. Approximately, it turns out that the chance of a Pandora appearing, assuming they are all allowed to spawn, is in these proportions:
Troops 42 (14 units of 3)\Experience 80\Gold 20\Magic level 10\Magic school 10
That is, in %, it would be 25/50/13/6/6?

In a certain sense, yes (if you consider all magic as magic of a school). But this does not take into account the fact that all Pandoras have different values - and because of this, some of them cannot be generated at all, or some of them have an advantage due to their value. Therefore, it makes more sense to compare them for specific value ranges. For example, if the total value should be 5000-6000, then this is usually a level 1 creature with or without an upgrade, 5000 experience, gold, or level 1 spells. Frequencies of 6/20/5/2/0, i.e., 19/60/15/6/0, are quite close to your average estimate. But there is also another point: for half of the objects in the zone, the generator tries to choose the most "dense" object, i.e., the one with the maximum value divided by the number of cells. In the case of Pandoras, for example, in this situation, a more valuable Pandora with experience will displace a Pandora with gold, i.e., the comparison of frequencies will not even occur.