fix(model_metadata): parse Google's 'supports up to N' context-limit phrasing

Google Gemini/Gemma overflow errors read 'Unable to submit request because
the input token count is 32825 but model only supports up to 32768'.
parse_context_limit_from_error had no pattern for the 'supports up to N'
phrasing, so overflow recovery kept the wrong window and burned its retry
attempts instead of recalibrating to the provider-reported limit.

Add the anchored pattern (limit follows 'supports up to'; the larger input
count before it is never captured) plus regression tests covering the exact
message and the get_context_length_from_provider_error recalibration path.

Reported by @Artemonim in #57275 (residual claim 5).

agent/model_metadata.py

@@ -1679,6 +1679,7 @@ def parse_context_limit_from_error(error_msg: str) -> Optional[int]:
       - "context_length_exceeded: 131072"
       - "Maximum context size 32768 exceeded"
       - "model's max context length is 65536"
+      - "input token count is 32825 but model only supports up to 32768"
     """
     error_lower = error_msg.lower()
     # Pattern: look for numbers near context-related keywords
@@ -1690,6 +1691,12 @@ def parse_context_limit_from_error(error_msg: str) -> Optional[int]:
         r'(\d{4,})\s*(?:token)?\s*(?:context|limit)',
         r'>\s*(\d{4,})\s*(?:max|limit|token)',  # "250000 tokens > 200000 maximum"
         r'(\d{4,})\s*(?:max(?:imum)?)\b',  # "200000 maximum"
+        # Google Gemini/Gemma: "Unable to submit request because the input
+        # token count is 32825 but model only supports up to 32768." The
+        # limit is the number AFTER "supports up to" — the input count that
+        # precedes it must not be captured, so this pattern anchors on the
+        # "supports up to" phrase itself.
+        r'supports?\s+(?:only\s+)?up\s+to\s+(\d{4,})',
     ]
     for pattern in patterns:
         match = re.search(pattern, error_lower)

tests/agent/test_model_metadata.py

@@ -1413,6 +1413,29 @@ class TestParseContextLimitFromError:
         fell through to None."""
         assert parse_context_limit_from_error(msg) == expected
 
+    @pytest.mark.parametrize("msg,expected", [
+        # Google Gemini/Gemma overflow phrasing (#57275): the limit follows
+        # "supports up to"; the larger input count before it must NOT win.
+        ("Unable to submit request because the input token count is 32825 "
+         "but model only supports up to 32768. Reduce the input token count "
+         "and try again.", 32768),
+        ("input token count is 140000 but model only supports up to 131072", 131072),
+        ("model supports up to 65536 tokens", 65536),
+    ])
+    def test_google_supports_up_to_variants(self, msg, expected):
+        """Google's overflow error was previously unparseable — recovery kept
+        the wrong window and burned its attempts (#57275, residual claim 5)."""
+        assert parse_context_limit_from_error(msg) == expected
+
+    def test_google_supports_up_to_recalibrates_window(self):
+        from agent.model_metadata import get_context_length_from_provider_error
+
+        msg = ("Unable to submit request because the input token count is "
+               "32825 but model only supports up to 32768.")
+        assert get_context_length_from_provider_error(msg, 131072) == 32768
+        # Parsed limit not below current window → no recalibration.
+        assert get_context_length_from_provider_error(msg, 32768) is None
+
     def test_get_context_length_from_vllm_max_model_len_error(self):
         from agent.model_metadata import get_context_length_from_provider_error