Solving Sudoku with TDD

I ran across the Ravi Mohan post last year that compares Jeffries' and Norvig's approaches to solving Sudoku. Upon reading the actual articles, I felt that the comments in that thread were somewhat unfair.

Jeffries explicitly states that he is unfamiliar with Sudoku and rusty with Ruby, so of course he is going to make mistakes as he seeks a solution; it doesn't help his perceived credibility that he makes lots of little mistakes and documents his fixes for each one. His Test-Driven Design approach makes for a long and boring read, but it is what enabled him to make as much progress as he did.

Norvig, on the other hand, has a masterful command of Python and has probably worked through dozens of Sudoku by hand, never mind that he is intimately familiar with constraint propagation and doesn't bother describing any problems he had typing up the program.

To Jeffries' credit, he seems to hit upon the same solution as Norvig in the very first article (a "tree-trimming" solution, as he puts it), but he doesn't get around to implementing it. He states that he doesn't think it will run fast enough for all puzzles -- a perfectly reasonable concern that Norvig ducks entirely by running his solution on a hundred puzzles and timing them, then assuming they constitute the hardest puzzles the algorithm will encounter, in the astronomical space of all puzzles. It's not an unreasonable assumption, but it is an assumption nonetheless; there could well be a Sudoku puzzle out there that causes Norvig's algorithm to run an unreasonably long time.