Look at the following document:

  \Umathlimitbelowbgap\displaystyle0pt
  \Umathlimitbelowvgap\displaystyle0pt
  $$
    \sum_a^b
  $$
  \bye

Here the `a` subscript gets lost. In mlist.c, make_op, line 3324 the
subscript (stored in variable `z`) is never coupled with the previous
node if `shift_down` is smaller or equal to 0. This can be fixed by

---
 source/texk/web2c/luatexdir/tex/mlist.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/source/texk/web2c/luatexdir/tex/mlist.c b/source/texk/web2c/luatexdir/tex/mlist.c
index 43db6adbb..570a54052 100644
--- a/source/texk/web2c/luatexdir/tex/mlist.c
+++ b/source/texk/web2c/luatexdir/tex/mlist.c
@@ -3321,7 +3321,9 @@ static scaled make_op(pointer q, int cur_style)
             shift_down = limit_below_bgap(cur_style) - height(z);
             if (shift_down < limit_below_vgap(cur_style))
                 shift_down = limit_below_vgap(cur_style);
-            if (shift_down > 0) {
+            if (shift_down == 0) {
+                couple_nodes(y, z);
+            } else {
                 p = new_kern(shift_down);
                 reset_attributes(p, node_attr(q));
                 couple_nodes(y,p);
--
2.26.1

This also allows shift_down to be negative. While a negative shift_down
is kind of odd, it is already allowed for the superscript so it makes
sense from a consistency point of view. Also it can't happen except when
it is explicitly requested though a negative limit_below_vgap, so it
doesn't have any effect if it isn't requested by the user.