fix(codex-bridge): estimate Claude requests more strictly; report an unexplained 400 on a near-full request

A 607 KB Codex request was estimated at 189776 tokens (window 200000), sent untrimmed and rejected by
Claude; the gateway only said "upstream 400", so Codex saw a dropped stream. Claude targets now use
2.6 bytes/token, and that bare 400 on a request filling most of the window becomes context_length_exceeded.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

apps/desktop/assets/codex-bridge/test_pencere_phase.py

@@ -225,5 +225,20 @@ class Images(unittest.TestCase):
         self.assertEqual(resim_context.resimleri_sadelestir(big, tut_bayt=500, gorsel=True)[0], 2)   # over budget
 
 
+
+class ClaudeOrani(unittest.TestCase):
+    def test_claude_targets_are_estimated_more_strictly_than_others(self):
+        import pencere as p
+        req = {"input": [{"type": "function_call_output", "call_id": "c", "output": "x" * 600_000},
+                         {"type": "message", "role": "user", "content": "son"}]}
+        self.assertEqual(p.oran_sec("9router-cc-claude-opus-5-5"), p.CLAUDE_BAYT_BOLU_TOKEN)
+        self.assertEqual(p.oran_sec("9router-ag-gemini-3.8-flash-low"), p.BAYT_BOLU_TOKEN)
+        self.assertEqual(p.sigdir({"input": [dict(i) for i in req["input"]]}, 200_000)[0], 0)          # 187500: fits
+        trimmed, fits, estimate = p.sigdir(req, 200_000, p.oran_sec("cc-claude-opus"))                 # 230769: trimmed
+        self.assertEqual((trimmed, fits), (1, True))
+        self.assertLess(estimate, 194_000)
+        self.assertTrue(p.kapali_hata_mi("upstream 400"))
+        self.assertFalse(p.kapali_hata_mi("upstream 400: invalid tool schema"))
+
 if __name__ == "__main__":
     unittest.main()