At this point I disagree. It’s faster to fetch one row using seq scan that using index scan as well as fetching number of consecutive rows is faster via seq scan. Index scan is not always faster.
Why planner mistakes in determining the number of rows (every time planner expects only 1 row) in this step I can understand - inner nodes do some joins (inner and outer with filtration) and it's hard to predict result. But what I can't understand is why seq scan when it is always slower than index.