fix missing loop check

We failed to detect cyclic dependencies, leading to lost thread
roots.  In this Ruby version, I also screwed up the `seen' hash
check in a misguided effort to avoid an extra hash lookup.

The same algorithm was recently fixed in public-inbox:

  https://public-inbox.org/meta/20180425085249.14974-1-e@80x24.org/

Summarizing what happened with public-inbox: this algorithm can
still be thrown off when the References: order presented to us
is wrong, so sorting messages by age before feeding to
Msgthr#add can improve the results.

Regardless of ordering, messages should not become "lost" by the
algorithm.

1 files changed, 30 insertions, 0 deletions

diff --git a/test/test_msgthr.rb b/test/test_msgthr.rb
index ee04d54..3d70d35 100644
--- a/test/test_msgthr.rb
+++ b/test/test_msgthr.rb
@@ -144,4 +144,34 @@ EOF
     assert_equal threads[2][0], '2'
   end
 
+  def test_no_lost_root
+    ids = [
+      [ 8, [] ],
+      [ 7, [8] ],
+      [ 6, [8, 7] ],
+      [ 3, [6, 7, 8] ],
+      [ 2, [6, 7, 8, 3] ],
+      [ 10, [8, 7, 6] ],
+      [ 9, [6, 3] ],
+      [ 5, [6, 7, 8, 3, 2] ],
+      [ 4, [2, 3] ],
+      [ 1, [2, 3, 4] ],
+      [ 'a', ],
+      [ 'b', ['a'] ],
+    ]
+    [ [ :forward, ids, 2 ],
+      [ :backwards, ids.reverse, 3 ],
+      [ :shuffle, ids.shuffle, nil ],
+    ].each do |desc,msgs,exp|
+      thr = Msgthr.new
+      msgs.each { |id| thr.add(id[0], id[1], id[0]) }
+      seen0 = nr = 0
+      thr.walk_thread do |level, _, _|
+        seen0 += 1 if level == 0
+        nr += 1
+      end
+      assert_equal nr, msgs.size, 'no lost messages'
+      assert_equal exp, seen0, "single root #{desc}" if exp
+    end
+  end
 end