P5-2 prior-sensitivity sweep

Each of 40 sparse persona × conflict-cell clusters receives an independent symmetric Dirichlet prior; at α=0.5 the total prior concentration (≈60 category-count units) is non-negligible against 352 observed episodes and pulls the aggregate toward 0.5. The sweep shows the Jeffreys posterior’s proximity to the registered 0.20 boundary is a prior artifact, not a data signal — the empirical anchor stays 45/352 = 0.128 and the prompt-cluster bootstrap [0.071, 0.189] remains the principal dependence-aware sensitivity. This correction (Round 14) demoted “independent Bayesian corroboration” from the P5-2 story.