While a few algorithm and coding bugs may be left in the code, the most common source of errors is incorrect input. Be sure to check the input parameters against the documentation and the USER_INPUT routine. The correctness of the input property file format can be found by using the 'M' - medium property output option. There are routines that check the input parameters and property file for validity and issue warnings for common mistakes.
Conservation of energy can be used as a check for conservatively scattering medium, but remember that the multiple reflections between the surface and medium will introduce "extra" flux. The collimated (solar) flux input parameter is the flux on a horizontal surface (rather than the flux perpendicular to the beam), so for conservative scattering and no surface reflection the reflected plus transmitted flux over the domain should equal this parameter.
The delta-M scaling procedure is very useful for reducing the angular resolution needed. It is generally used for collimated (solar) source problems. One should remember that the delta-M scaling procedure changes the extinction, single scattering albedo, and Legendre phase function fields, as can be seen with the medium output option. This means that the direct solar flux is increased with delta-M scaling and must be added to the diffuse transmitted flux for meaningful results.
For putting in the rest of the atmosphere: a coarse grid outside the cloud region works because the rest of the atmosphere does not interact much with the cloud layer (little reflection back to the cloud layer). The fluxes, which depend on the cell-to-cell discrete ordinate integration, will not be as reliable above the cloud level. The radiances, however, which are computed by integrating the source function back all the way to the boundary will be fine at any level.
The solution criterion often does not converge monotonically. When the cell splitting accuracy parameter is lowered and many adaptive cells are created the solution criterion typically increases. The sequence acceleration method may cause the solution criterion to increase at a particular step, though the overall convergence is usually speeded. The adaptive spherical harmonic truncation (parameter SHACC) may also interfer with a smooth convergence. If you have trouble with convergence try adusting one of these parameters (SPLITACC, SHACC, ACCEL).