Utility Non-Monotonicity with α < 1 (Stub)

@cite{zaslavsky-hu-levy-2020}

Formalizes Proposition 2 from @cite{zaslavsky-hu-levy-2020}:

Expected listener utility E[V_L] can decrease during RSA iterations, even though G_α increases monotonically.

This result distinguishes RSA from pure information-theoretic optimization: the objective G_α can improve while actual communication accuracy temporarily decreases.

Counterexample Requirements (from @cite{zaslavsky-hu-levy-2020}, page 3)

Three conditions are necessary for utility non-monotonicity:

1. Graded lexicon: φ(m,u) ∈ (0,1], not binary {0,1}.
2. α < 1: Below the rate-distortion critical point.
3. Non-maximal initial entropy: Start from (L₀, S₁), not uniform.

Old RSAScenarioQ-based counterexample removed; restate using RSAConfig.