I checked the alter table code, and started suspecting the relation

cache.  I found if I manually added a line to flush the whole relation
cache, the assert error disappeared.  Looking through the code, I found
that the relation cache is flushed at the end of each query if the
reference count is zero for the relation.  However, printf's showed that
the rd_relcnt(reference count) for the accessed query was not returning
to zero after each query.

It turns out the parser was doing a heap_ropen in parser/analyze.c to
get information about the table's columns, but was not doing a
heap_close.

This was causing the query after the ALTER TABLE ADD to see the old
table structure, and the executor's assert was reporting the problem.