_BTree_set(): Fixes for "Guido's bug", and all the other kinds of BTree

deletion endcases uncovered by the new degenerate-BTree tests.  The
degenerate testDeletes() and testEmptyFirstBucketReportedByGuido() are
enabled now.

src/BTrees/Maintainer.txt

@@ -173,8 +173,11 @@ More or less random bits of helpful info.
   more efficiently than they can store arbitrary Python objects.
 
 + In a non-empty BTree, every bucket node contains at least one key,
-  and every BTree node contains at least one child.  An empty BTree
-  consists solely of a BTree node with len==0 and firstbucket==NULL.
+  and every BTree node contains at least one child and a non-NULL
+  firstbucket pointer.  However, a BTree node may not contain any keys.
+
++ An empty BTree consists solely of a BTree node with len==0 and
+  firstbucket==NULL.
 
 + Although a BTree can become unbalanced under a mix of inserts and
   deletes (meaning both that there's nothing stronger that can be

src/BTrees/tests/testBTrees.py

@@ -759,7 +759,7 @@ class TestIOBTrees(BTreeTests, TestCase):
     def _noneraises(self):
         self.t[None] = 1
 
-    def XXXtestEmptyFirstBucketReportedByGuido(self):
+    def testEmptyFirstBucketReportedByGuido(self):
         b = self.t
         for i in xrange(29972): # reduce to 29971 and it works
             b[i] = i
@@ -994,20 +994,18 @@ class TestIITreeSets(NormalSetTests, TestCase):
         t, keys = self._build_degenerate_tree()
         self._checkRanges(t, keys)
 
-    def XXXtestDeletes(self):
+    def testDeletes(self):
         # Delete keys in all possible orders, checking each tree along
         # the way.
 
-        # XXX This is hopeless for now:  it dies with:
-        # XXX 1. A variety of assertion failures in _checkRanges.
-        # XXX 2. Assorted "Invalid firstbucket pointer" failures at
-        # XXX    seemingly random times, coming out of the BTree destructor.
-        # XXX 3. Under Python 2.3 CVS, some baffling
-        # XXX    RuntimeWarning: tp_compare didn't return -1 or -2 for exception
-        # XXX    warnings, possibly due to memory corruption after a BTree
-        # XXX    goes insane.
-        # XXX These are probably related to "Guido's bug" (which test case
-        # SXX is also disabled for now).
+        # This is a tough test.  Previous failure modes included:
+        # 1. A variety of assertion failures in _checkRanges.
+        # 2. Assorted "Invalid firstbucket pointer" failures at
+        #    seemingly random times, coming out of the BTree destructor.
+        # 3. Under Python 2.3 CVS, some baffling
+        #      RuntimeWarning: tp_compare didn't return -1 or -2 for exception
+        #    warnings, possibly due to memory corruption after a BTree
+        #    goes insane.
 
         t, keys = self._build_degenerate_tree()
         for oneperm in permutations(keys):
@@ -1016,6 +1014,9 @@ class TestIITreeSets(NormalSetTests, TestCase):
                 t.remove(key)
                 keys.remove(key)
                 self._checkRanges(t, keys)
+            # We removed all the keys, so the tree should be empty now.
+            self.assertEqual(t.__getstate__(), None)
+
             # A damaged tree may trigger an "invalid firstbucket pointer"
             # failure at the time its destructor is invoked.  Try to force
             # that to happen now, so it doesn't look like a baffling failure