RECOMMENDATIONS FROM THE STOCHASTIC AND EXPERIMENT
==================================================

1. `Mind::step` must execute declared impure expressions in `neuron.output`,
   especially random(), normal(), bernoulli(), and poisson(), or the parser must
   reject constructs that the selected runtime cannot execute.

2. The result of `fire when` must be observable. Suggested REST output:

       { "activation": 0.42, "fired": true, "fire_probability": 0.73 }

3. `on fire` assignments must execute in the same tick and their effects must be
   visible through /step, /activation, and actuators.

4. /reset must document whether it also resets per-neuron RNG state. /evaluate
   should optionally support `reset_rng: false`; otherwise repeated stateless
   requests replay the same sample in a seeded model.

5. Add a regression test with two sensors and one neuron using
   `fire when random() < sigmoid(k*(net-threshold))`. Over 10,000 stateful steps,
   empirical frequencies must match theory within tolerance, and resetting with
   the same seed must reproduce the sequence.
