You paint an accurate picture of the current state of affairs where we’re unable to make improvements without incurring massive costs for each change. But surely, the conclusion cannot be to stop improving the OSM data model. Even if we can live without changes that only affect people looking at raw tags (which includes data consumers who may consider adopting OSM in the future, and for whom an idiosyncratic and messy data model is a turn-off), all these same costs are incurred by, and therefore deter us from, making changes that do introduce new capabilities.
Therefore, my conclusion is that we need a better process for making changes to the OSM data model. This might very well involve a versioned specification for at least the “core” of the OSM tagging data model, which I expect would make it both more feasible to make changes to the data model and make tracking those changes as a data consumer less onerous. (After all, what’s the current alternative? Watch a few thousand wiki pages and sift out the substantial changes from the formatting tweaks?)
Of course, pulling that off would require the OSM community to demonstrate an uncharacteristically high level of coordination.