I tried an even more simple test (create table dummy (a varchar2(2))), and was not able to reproduce the problem. Still investigating, but I am very suspicious. Another anomaly is that running my original sql with a 10053 did not produce a trace file (statistics_level = all). The last, very simple test did, however.
I filed a bug with Oracle. Apparently, bind buffers are not peeked when you set statistics_level = ALL in 10.2.0.2. At least, not for my simple tests.
I was using statistics_level = all because it makes Cardinality Feedback so much simpler (query 2 dictionary tables, as opposed to ripping apart 10046 and 10053 traces). So much for that idea....
-- Charles Schultz