> Is it normal that even after the call to next(), the generator frame 
> contains a reference to the RuntimeError?

As long as it's executing the generator it's quite normal. The generator must be able to restore the caller's RuntimeError when interrupted by a yield. But after the yield, the generator normally doesn't retain a reference (thanks to SWAP_EXC_STATE() in YIELD_VALUE).

> - This executes SWAP_EXC_STATE() in ceval.c ("/* We were in an except handler when we left, restore the exception state which was put aside */"), this *moves* the reference to the exception from the frame to the thread state.

Actually, it (does/should) execute SAVE_EXC_STATE() instead, because the thread's exception state was empty when yielding.

The following patch seems to do the trick:

diff -r 4465a45b8876 Python/ceval.c
--- a/Python/ceval.c	Mon Feb 15 09:35:16 2010 +0100
+++ b/Python/ceval.c	Mon Feb 15 20:51:58 2010 +0100
@@ -1159,7 +1159,7 @@ PyEval_EvalFrameEx(PyFrameObject *f, int
 	assert(stack_pointer != NULL);
 	f->f_stacktop = NULL;	/* remains NULL unless yield suspends frame */
 
-	if (f->f_code->co_flags & CO_GENERATOR) {
+	if (!throwflag && (f->f_code->co_flags & CO_GENERATOR)) {
 		if (f->f_exc_type != NULL && f->f_exc_type != Py_None) {
 			/* We were in an except handler when we left,
 			   restore the exception state which was put aside