Unknown 'species means'. Allometry can have different effects on different timescales, generating patterns like Simpson's Paradox. Three processes in particular are likely:

• universal physical constraints can be modeled with logmass directly as predictor
• evolutionary optima can be modeled by using a species' mean size as the predictor rather than the individual itself (e.g. the response variable is not directly modulated by size, but selection has pulled it toward an optimum based on the size of recent ancestors. The evolved mechanisms to produce the response phenotype might not be 'aware' of an individual's deviance from its ancestors' average size, so the phenotype is more correlated with the ancestors than with the individual)
• deviance from the optimum can have direct effects; either absolute deviance or directional. Anna Karenina-type effects or incidental correlation with ontogenetic development, for instance

A basic linear model can contain e.g. two of these effects directly in order to control for the other and resolve simpson's paradox. But species are not discrete, mean values are only estimates, and the lag-time between a phenotype evolving toward a population's optimum is unknown. Thus it may make sense to not pre-calculate species means and deviances, but rather to model them simultaneously with other parameters. Practically, species means should be shrunken toward the means of relatives. Multiple 'reference means' could potentially be modeled simultaneously, representing different shrinkage functions that correspond to different scales. Thus you may have one predictor that has practically no shrinkage at all, corresponding to a universal effect, and another that has a lot of local shrinkage, corresponding to something more like discrete species means. These will likely be highly correlated and thus difficult to tell apart, but post-processing may be able to disentangle possibilities.