ALL 72 units: margin +0.0511  accuracy 0.5972

--- design-controlled covariates (constant within every eligible receptor) ---
  experimental_method  varies within receptor in 0/34 receptors
  species              varies within receptor in 0/34 receptors
  ligand_type          varies within receptor in 1/34 receptors

--- A. stratification by class pair (stabiliser proxy) ---
  Gi/o+Gq/11           receptors 18 units 36  margin +0.0850 (perm -0.0003, P=0.0380)
  Gi/o+Gq/11+Gs        receptors  4 units 12  margin -0.1378 (perm -0.2143, P=0.3093)
  Gi/o+Gs              receptors  6 units 12  margin +0.0885 (perm +0.0017, P=0.2092)
  Gq/11+Gs             receptors  6 units 12  margin +0.1012 (perm -0.0006, P=0.0600)

--- B. same-publication subset (same lab, same paper, same construct strategy) ---
  receptors 8 units 16  margin +0.0973 (perm -0.0001, P=0.1719)
  receptors: 5ht4r_human, gasr_human, hrh2_human, kissr_human, mrgx1_human, pe2r4_human, prlhr_human, us28_hcmva

--- C. covariate adjustment (within-receptor centred) ---
  within-receptor covariates explain R^2 = 0.0212 of the margin variance
  coefficients: resolution +0.0976  year +0.09391  fusion -0.1005  ICL2-ligand -0.0010
  adjusted mean margin +0.0511 (unadjusted +0.0511)

--- D. resolution-matched subset (|delta resolution| <= 0.3 A within receptor) ---
  receptors 19 units 40  margin +0.0736 (perm -0.0373, P=0.0329)
