Seed granularity (shared vs per-source) is a design choice between measuring total noise and attributing it to sources

Having multiple components accepting a random_seed, we can either:

• Provide a single, shared seed as input to all seed-accepting functions. We can compare total noise across pipeline runs.
• Provide a per-source seed as inputs. This allows us to run each of them true ceteris paribus - by flipping a single seed at a time.