The .osrm file contains a node-based graph representation of the OSM data, prior to compression and edge-expansion.

It's generated by osrm-extract and only read by osrm-extract. I believe it's a hangover from when the data was too large for RAM, so was first serialised to disk. However the file contents are immediately read back into memory anyway.

Assuming the file had value beyond the extraction phase, there are inconsistencies in which metadata components are serialised.
Traffic signals and barriers are serialised to disk as part of the graph, but other metadata, such as turn lanes, restrictions and manoeuvrer overrides are not. They are only kept in memory long enough for subsequent extraction steps.

Therefore, I propose that we scrap the disk serialise/deserialise step and stop generating this file.

.osrm is one of, if not the largest file generated during processing. Given we generate many other specific files for use in later processing stages, we can stop persisting the .osrm file entirely to save storage space, and if needed serialise its contents to more targeted files.