This seems beyond me at this point, but I am curious if you also vacuumed alf_node_properties and alf_node tables and checked when they last got (auto)vacuumed/analyzed. With default configs for autovacuum parameters and tables with that many rows, they don't qualify for autovacuum very often. I don't have much experience with tables in excess of 50 million rows because of manual sharding clients data.
You mention work_mem is set differently. Did you try setting work_mem back to 4MB in session on instance 1 just to test the query? I don't know if work_mem is included in planning stage, but I would think it may be considered. It would be odd for more available memory to end up with a slower plan, but I like to eliminate variables whenever possible. It might be worthwhile to see about increasing default_statistics_target to get more specific stats, but that can result in a dramatic increase in planning time for even simple queries. Hopefully one of the real experts chimes in.
