v0.3.2

What's Changed

• Alpha + weights composability in log_odds_conjunction()

  • Per-signal weights (Theorem 8.3) and confidence scaling by signal count (Section 4.2) compose multiplicatively: sigma(n^alpha * sum(w_i * logit(P_i)))
  • alpha default changed from 0.5 to None for backward compatibility: resolves to 0.5 in unweighted mode and 0.0 in weighted mode
  • Explicit alpha applies in both unweighted and weighted modes

• 24 new theorem verification tests covering:

  • Strict bounds for prob_and / prob_or (Theorem 5.1.2 / 5.2.2)
  • Log-OP / Product of Experts equivalence (Theorem 4.1.2a)
  • Heterogeneous signal combination -- BM25 + cosine pipeline (Remark 5.2.3)
  • Single signal identity for all alpha values (Proposition 4.3.2)
  • Weighted alpha composition (Section 4.2 + Theorem 8.3)
  • Monotone shrinkage of product rule (Theorem 3.2.1 + Corollary 3.2.2)
  • Information loss in product rule vs log-odds conjunction (Proposition 3.4.1)
  • sqrt(n) scaling law (Theorem 4.4.1 + Proposition 4.4.2)
  • Spread property -- disagreement reduces confidence (Theorem 4.5.1 (iii))
  • Geometric mean residual vs log-odds mean (Remark 4.1.3)
  • Sigmoid uniqueness -- alternative activations fail (Theorem 6.2.1)

Full Changelog: v0.3.1...v0.3.2