Improve random() to use name-based salting for order-independent reproducibility

[baogorek]

Problem

The current random() function in policyengine_core/commons/formulas.py uses a global execution counter (count_random_calls) to differentiate random streams:

seeds = np.abs(entity_ids * 100 + population.simulation.count_random_calls)

This creates a "ripple effect": adding, removing, or reordering variables that call random() changes the random values for ALL subsequent variables. This makes it impossible to:

• Compare policy versions with confidence (random noise shifts underneath)
• Isolate the effect of a specific policy change
• Run variables in parallel without counter synchronization

Proposed Solution: Name-Based Salting

Replace the global counter with the variable name (accessible via population.simulation.tracer.stack[-1]["name"]):

base_seed = stable_hash(f"{variable_name}:{per_variable_call_count}")
seeds = entity_ids ^ base_seed

Benefits:

• Order-independent: Adding/removing variables doesn't affect others
• True reproducibility: Same variable + entity ID = same value, always
• Parallelizable: No global state to synchronize

Breaking Change

This will change random values for all existing simulations using random(). Downstream packages (policyengine-us, policyengine-uk) will see different takeup modeling results.

Questions for Maintainers

1. Is this change acceptable given the breaking nature?
2. Should we provide a legacy_random() for transition?
3. Any concerns about the tracer stack approach?