The tools use ffmpeg to stream process the videos, and it has a concatenate functionality if useful. I advise some caution though as (in the BlackVue device I use) there is a frames/time overlap between movies that would result in a timing (and GPS) error getting worse with each concat. The tools also use ffprobe to extract some data (start time, rate and total frames perhaps), so that’s a concern (ffmpeg can be set to drop frames intentionally though.)
The BlackVue device I use creates a .gps and .mp4 file(name) that are correlated, so I thought I’d try your (failed) video_process method. I took 3 mp4’s against converted gpx for the whole day, but it came back with full geotagging errors. It didnt crash like yours though. I suspect that in this mode it falls back to the EXIF tag “Original Date and Time”, which comes from the file date. I grab my mp4’s over WiFi with wget and this seems to result in a 63 second error. (ie the difference between creation and modified dates) If I continue the experiment I’ll likely force the video start with --video_start_time (epoch seconds)