Skip the decode round-trip check for domains past encode's length limit

tests/fuzz_idna_api.py

@@ -62,14 +62,16 @@ def fuzz_decode(fdp):
         return
     # RFC 5891 §5.3: an A-label that decodes must re-encode to itself, so
     # under strict non-UTS46 processing any ASCII input that decodes is (up
-    # to case) its own encoding. decode() does not enforce the 63-octet
-    # label limit (nor does UTS #46 ToUnicode) but encode() does, so skip
-    # overlong labels.
+    # to case) its own encoding. decode()'s length checks are deliberately
+    # lenient (no 63-octet label limit, and the whole-domain limit allows
+    # the trailing-dot octet whether or not one is present, as UTS #46
+    # ToUnicode checks no lengths at all) while encode() enforces both
+    # exactly, so skip inputs past either limit.
     if isinstance(data, str):
         if not data.isascii():
             return
         data = data.encode("ascii")
-    if all(len(label) <= 63 for label in data.split(b".")):
+    if idna.valid_string_length(data, data.endswith(b".")) and all(len(label) <= 63 for label in data.split(b".")):
         assert idna.encode(decoded, strict=True) == data.lower(), (data, decoded)
 
 

tests/test_idna.py

@@ -79,6 +79,18 @@ def test_valid_label_length(self):
         self.assertFalse(idna.valid_label_length("a" * 64))
         self.assertRaises(idna.IDNAError, idna.encode, "a" * 64)
 
+    def test_decode_domain_length_leniency(self):
+        # UTS #46 ToUnicode checks no lengths, so decode()'s whole-domain
+        # bound is a lenient pre-filter that always allows the trailing-dot
+        # octet, while encode() enforces the exact RFC 1035 limit: a
+        # 254-octet domain without a trailing dot decodes but does not
+        # re-encode (OSS-Fuzz 4694619275984896), and with one it round-trips.
+        domain = b"aaa." * 63 + b"aa"  # 254 octets, every label within limits
+        decoded = idna.decode(domain, strict=True)
+        self.assertRaises(idna.IDNAError, idna.encode, decoded, strict=True)
+        dotted = domain[:-1] + b"."
+        self.assertEqual(idna.encode(idna.decode(dotted, strict=True), strict=True), dotted)
+
     def test_oversized_input_rejected_promptly(self):
         # GHSA-65pc-fj4g-8rjx: encode/decode must reject inputs that
         # exceed the maximum DNS domain length before per-codepoint

tests/test_idna_fuzz_targets.py

@@ -43,6 +43,7 @@
     b"xn---bbk.example",  # non-canonical spelling of xn--bbk
     b"XN--MNCHEN-3YA.example",
     b"a." * 200,  # every label valid, but past the 253-octet domain limit
+    b"aaa." * 63 + b"aa",  # 254 octets, no trailing dot: decodes, too long to encode
     ("\ufdfa" * 100).encode(),  # UTS #46 maps each to 18 characters, past the input cap
 ]