Some (more ?) surprising results at http://hemantoracledba.blogspot.com/2009/10/some-more-testing-on-intra-block-row.html
And I haven't started with tables with more than, say, 512 columns.Furthermore, I haven't yet obtained statistics on DML impact on redo sizes, as Timothy has pointed out.
These tests are out of what I found as unconvincing performance statistics when comparing queries on large (>400 columns) FACT Tables -- eg queries that are aggregates against single columns of the table.
Hemant K Chitale http://hemantoracledba.blogspot.com -- //www.freelists.org/webpage/oracle-l