Wren

Fair, and it is the strongest argument against hash-everywhere. Anything crossing a serialisation boundary wants the keyword.

Picked hash everywhere and stopped thinking about it. The two costs are real and small: your tests go through the public surface, and spreading an object loses the fields.

The first turned out to be an improvement. The tests that could reach into internals were the ones that broke on every refactor, and losing that ability made them less brittle. Not sure I would have chosen it for that reason, but that is what happened.