Flow accumulation from r.watershed creates discontinuous stream network compared with ArcGIS

I am interested in using r.watershed as integrated in QGIS for a course I teach, but am noticing discrepancies in the flow accumulation (facc) rasters and resulting stream networks that I get from QGIS-GRASS (r.watershed) compared with ArcGIS. The r.watershed stream network is discontinuous (see blue-green cells below), compared with the continuous network of ArcGIS (red cells below; red layer plotted beneath the blue one). The facc grids were generated using the same DEM filled with ArcGIS, and both use a threshold of 500 cells for streams. The facc values from QGIS-GRASS also show high values surrounded by lower values at the discontinuities, indicating what should be a sink, but ArcGIS does not see them as sinks on this filled DEM, and filling the DEM again does not change the result. I am using QGIS V 3.12.2 (see image below for full version information).

I have seen other answers to filling in a discontinuous stream network (e.g. here), but don’t agree that you should have to paste over these gaps with any further raster-vector processing, since facc values should always decrease going downstream.
I also find that the r.water.outlet function on the same facc grid does not give a correct watershed boundary for some pour points (but does fine on others).

Is there an easy solution, and/or is this discontinuity in r.watershed a common issue?
Thank you!

The original DEM and the resulting facc rasters can be found here:
https://drive.google.com/open?id=17_RPXoyR7IVTo7K6UQ_92yWEAMny6a0i&authuser=tbiggs%40sdsu.edu&usp=drive_fs