test: assert the max_tokens error names the SDK's own number

tests/unit/test_validation.py

@@ -2,6 +2,7 @@
 
 import pytest
 from blockrun_llm.validation import (
+    MAX_TOKENS_SANITY_LIMIT,
     validate_private_key,
     validate_api_url,
     validate_model,
@@ -130,10 +131,6 @@ def test_accept_valid_values(self):
         validate_max_tokens(100)
         validate_max_tokens(1000)
         validate_max_tokens(100000)
-        # Real ceilings the gateway serves — these were rejected before the
-        # sanity bound was raised, despite every provider accepting them.
-        validate_max_tokens(128000)  # opus-4.8 / sonnet-5 / gpt-5.6 / glm-5
-        validate_max_tokens(262144)  # zai/glm-5.2
 
     def test_accept_none(self):
         """Should accept None."""
@@ -151,17 +148,30 @@ def test_reject_zero(self):
 
     def test_reject_implausible(self):
         """Should reject values no model could mean (typo guard, not a limit)."""
+        with pytest.raises(ValueError, match="implausibly large") as exc:
+            validate_max_tokens(MAX_TOKENS_SANITY_LIMIT * 2)
+        # The message must name the SDK's own number, so a caller can tell it
+        # apart from a provider ceiling.
+        assert str(MAX_TOKENS_SANITY_LIMIT) in str(exc.value)
+
+    def test_boundary_is_inclusive(self):
+        """Pin the exact edge: the limit passes, one above it fails.
+
+        Without this, flipping ``>`` to ``>=`` keeps the suite green while
+        rejecting a legal value.
+        """
+        validate_max_tokens(MAX_TOKENS_SANITY_LIMIT)
         with pytest.raises(ValueError, match="implausibly large"):
-            validate_max_tokens(2_000_000)
+            validate_max_tokens(MAX_TOKENS_SANITY_LIMIT + 1)
 
     def test_does_not_cap_below_real_ceilings(self):
         """The bound must never be the binding constraint on a real request.
 
         Regression: this was 100000, which rejected every ceiling above it
         client-side — the caller saw a ValueError naming a limit no provider
-        had set, and the request never reached the network. Probed against the
-        live gateway 2026-07-21: 19 models advertise more than 100000 and all
-        19 accepted their advertised ceiling.
+        had set, and the request never reached the network. 128000 is the
+        common ceiling (opus-4.8 / sonnet-5 / gpt-5.6 / glm-5); 262144 is
+        zai/glm-5.2, the highest any model serves.
         """
         for real_ceiling in (128_000, 262_144):
             validate_max_tokens(real_ceiling)