Two conflicts, both where the compare study meets the random-field seeding. `seed` is async now (it draws a randnfun3 table first), so main's `seedWith` becomes async too and takes the table alongside the noise. That is what the study needs anyway: for a randnfun3 model the table *is* the initial condition, and it must be drawn once rather than per session — each session's bounding box comes from grid samples of the surface, so at different lmax the draws are near neighbours instead of one field. `sharedModes` draws it from the reference and every variant seeds from it (8x closer across lmax than per-session draws). The geometry note keeps main's compare-aware reference session and drops the "rendered only" caveat, which this branch retired by putting the surface in the operator. Also carried the seed wavelength into compare mode: it is passed to every variant, and the control now sets it on the study when one is up, rather than doing nothing.