Let's see below simplified file.
cluster ( #1, offset: 100, next_cluster_offset: 200 )
cluster ( #2, offset: 200, next_cluster_offset: 300 )
cluster ( #3, offset: 300, next_cluster_offset: 400 )
	............

cluster ( #100, offset: 10000, next_cluster_offset: 10100 )
  Unknown ID ( 0x00 )
  Unknown ID ( 0x00 )
  Unknown ID ( 0x00 )
	............

Demux is parsing currently #1 cluster, assuming seek to #100 cluster.
After handling seek event, demux->common.offset is 10000 and demux->next_cluster_offset is still 200.
Becasue the data is broken at 10000 offset, demux->common.offset will be changed to 200 in gst_matroska_demux_check_parse_error() function.
Demux will parse next cluster from 200 offset, but demux should search next cluster from 10000 offset.
demux->next_cluster_offset should be reset to 0 before the end of handling seek event.
So, I think it is a kind of solution to add below 2 lines in gst_matroska_demux_move_to_entry() function.

  demux->common.offset = entry->pos + demux->common.ebml_segment_start;
  /* reset next_cluster_offset */
  if( reset )
    demux->next_cluster_offset = 0;