Submerged jet impingement is commonly used in drying, cleaning and cooling applications. Recent theoretical work by our group has predicted all aspects of the flow and subsequently the heat transfer, including even the non-monotonous occurrence of an off-center heat transfer peak, as compared against experiments and simulations. This prediction is viable over a wide range of flow rates (200<Re<2000), fluids (Pr>0.1) and pipe-type geometries: dimensionless nozzle lengths (l/d) and jet-to-plate spacings (z/d) from around two to several hundred. Two limitations to this predictability are here numerically examined: first, the onset of jet-edge vortices that disrupt the smooth open-streamlines; and second, excessive proximity to the impinged plate that distorts the expected issuing profile. In the first case it is seen that the classic instability condition on profile curvature is necessary but insufficient, as roll-up only occurs above a critical inertia level (Re_c). Thereby a stable oscillating jet width regime is found in which vortex roll-up does not occur (Re< Re_c), where heat transfer is still well predicted. Roll-up has been shown to generate a secondary peak (around r~2d) in the heat transfer, which cannot be predicted by present theory. A regime map showing the onset of roll-up is presented, which is valid for all pipe-flow issuing profiles. Regarding the second limit, a new dominant flow-scale is identified, as being the distance at which the jet senses the presence of the plate (stagnation zone height, z_w). This new characteristic scale is shown to define the limit of excessive proximity to the impinged plate, rather than the geometric nozzle diameter (d). Therein it is found that only at much higher proximities (z<z_w/2), a sufficient back-pressure begins in the nozzle so as to modify the issuing profile. In other words, at intermediate proximities (2 z_w <z< z_w/2) a predictable, transition, self-similar impingement regime exists. Analysis of these two limits not only define the edges of current heat transfer predictability, but also show that they are much larger than previously thought, covering the vast majority of the relevant parameter range.