Implement O((M+N) log(M+N)) duplicate key detection in batchinsert, rather than naive O(N^2). Note that I haven't measured the constants in those complexity expressions; I am anticipating that 40000 is a sufficiently large N for this to matter. Write a test case for duplicate keys, too. Use 0037, since no-one else seems to be writing tests, and everything is merged onto the trunk these days anyway.