In my experience working with frequency filters often very promising strategy fails miserably on WFO stage.
However when I switch long and short entries after training,
or train with detrend=8, and test with detrend=0, the same WFO test becomes profitable.
Is there any theory behind this in your experience?
And can it be relied upon to trade?

Switching long and short trades is equivalent to going from a trend following to a trend anticipating system. Both can be profitable (or not) dependent on the details of the algorithm.

Training with detrend = 8 can indeed improve performance, especially when WFO periods are short and trends have different directions in the training and in the test period.