Out of interest, can you leave your memory monitoring hooks in place but test this ?
1) Create the table, allowing SQLite to assign ROWIDs, and without defining any indexes.
2) Load your 77 million rows of data, using transactions to bunch together INSERTs as before
3) Create a UNIQUE INDEX on <code>('col1','col2','col3')</code>
Does this turn out faster or slower than the procedure you describe above ? Does it allocate more or less memory ?