I beleive the problem is with ArcGIS
"HEC-GeoHMS uses ArcGIS and the Spatial Analyst extension to develop a number of hydrologic modeling inputs for the Hydrologic Engineering Center's Hydrologic Modeling System, HEC-HMS. ArcGIS and its Spatial Analyst extension are available from the Environmental Systems Research Institute, Inc. (ESRI)."
I don't use HEC-GeoHMS but I do use the ArcGIS and Spatial Analyst tools and I have seen these and other discrepancies over and over again especially when using but not limited to high resolution raster input (~ 1 - 5m resolution such as LiDAR)
ArcGIS uses an old algorytm that my be failing under certain conditions. Others have suggested to use more advanced algorytms such as those in GRASS GIS. In fact her is a quote from the GRASS Manual regarding its r.watershed function: "r.watershed will generate better basin results than r.terraflow. (r.terraflow uses the same type of algorithm as ESRI's ArcGIS watershed software which fails under these conditions.)"
You can read about some of my issues with hydrologic anlysisis in ArcGIS using high resolution data here: Methodology for creating accurate drainage networks (and catchments) from high resolution LiDAR DEM?
The answer by @david_p in the above linked post is also what I would advise to focus on...
I would strongly recommend using GRASS GIS which has a great Hydrologic Modelling tool set. It may be a little intimidating and there is a learning curve in even figuring out how to use the gui but the results from the Hydrolocic Tools, especially when used with high resolution raster input, are worth the effort. Another annoying fact is that GRASS uses proprietary raster and vector layer formats but those can be fairly easily exported to Shapefile or other spatial data types.