The separation of cosmological signal from astrophysical foregrounds is a fundamental challenge for any effort to probe the evolution of neutral hydrogen during the Cosmic Dawn and epoch of reionization (EoR) using the 21 cm hyperfine transition. Foreground separation is made possible by their intrinsic spectral smoothness, making them distinguishable from spectrally complex cosmological signal even though they are ~5 orders of magnitude brighter. Precisely calibrated radio interferometers are essential to maintaining the smoothness and thus separability of the foregrounds. One powerful calibration strategy is to use redundant measurements between pairs of antennas with the same physical separation in order to solve for each antennas spectral response without reference to a sky model. This strategy is being employed by the Hydrogen Epoch of Reionization Array (HERA), a large radio telescope in South Africa that is now observing while being built out to 350 14-m dishes. However, the deviations from perfect redundancy inherent in any real radio telescope complicate the calibration problem. Using simulations of HERA, we show how calibration with antenna-to-antenna variations in dish construction and placement generally lead to spectral structure in otherwise smooth foregrounds that significantly reduces the number of cosmological modes available to a 21 cm measurement. However, we also show that this effect can be largely eliminated by a modified redundant-baseline calibration strategy that relies predominantly on short baselines.