The thing worth noticing is that your markers were found but the molecule call wasn't. Those are two different steps in mob_recon: replicon, relaxase and mpf come from separate marker searches, while molecule_type comes from whether the contig gets assigned to a plasmid cluster by mash against the bundled reference database. Markers present plus a chromosome call means the clustering step failed, not the marker step, so min_length was never going to help you there.

Have a look at primary_cluster_id and mash_nearest_neighbor / mash_neighbor_distance in that same contig_report. If the cluster id is empty and the mash distance sits above threshold, then either nothing in the bundled DB was close enough (which is your database question, and it's a real possibility for a divergent plasmid) or the contig isn't purely plasmid.

Which brings up the other thing. You say the plasmid is 45 kb but the molecule is 36,253. That 9 kb gap seems worth chasing before touching any parameters. Either the remainder is sitting on other contigs, or the contig is a plasmid-chromosome chimera, which short-read assemblers do quite happily through shared IS elements, and IncX3 plasmids are IS-rich. If it is chimeric then the chromosome call is arguably correct rather than a bug. Two cheap checks: compare that contig's read depth against your chromosomal contigs, because a 45 kb plasmid sitting at the same coverage as the chromosome is suspicious, and blast the contig against a known IncX3 reference to see where the alignment stops.

If the assembly came out of Unicycler there's a --unicycler_contigs option that lets mob_recon use the circularity information in the headers, which fixes a fair number of cases like this. Long reads would settle it outright if you have any.