ca2dce9b48
Instead of collecting all potential inputs to some metadata entry and hashing those, we directly hash the values we are storing in metadata. This is more accurate and doesn't suffer from quadratic blow-up when many entries have the same dependencies.