The point about crossovers arriving late on M1 is easy to see once you measure it. By definition the fast average has to catch up with price and then pass the slow one, so a meaningful part of the move has already happened. With scalping targets that are only a few pips, the lag eats a large share of what's left.
The higher-timeframe filter is probably the most defensible use. Something like "only take longs when price is above the 50-period average on M15" doesn't try to time entries, it just keeps trades aligned with the broader direction. Whether it helps is testable: run the same entry rules with and without the filter and compare.
As dynamic levels, averages are a bit more debatable. They work partly because so many people watch them, and they're never as precise as a horizontal level, since the value changes with every bar. On M1 a 20-period average moves enough that "a pullback to the average" can mean quite different prices depending on how fast the pullback happens.
One other use worth mentioning is the distance between price and the average as a measure of stretch. When price is unusually far from a short-term average on M1, the move is extended, and that can be a reason not to chase it, even if the direction looks right. It's not a reversal signal, but it's a reasonable filter against entering at the worst point of a spike.
On testing crossovers with realistic costs, I agree it's a good exercise. It tends to show that the gross edge, if any, is small enough that spread and slippage remove most of it.