refactor(models): remove PR labels from runtime errors

sieval/cli/leaderboard/session.py

@@ -832,7 +832,7 @@ def reconcile(
                 CheckStage.REQUEST,
                 verifier,
                 requirement.reason
-                or "PR-1 compatibility requires response-time verification",
+                or "legacy compatibility requires response-time verification",
             )
         return outcomes
 
@@ -3301,7 +3301,7 @@ def replace_inline(
                     raise ValueError(
                         f"Model binding '{binding_id}' uses the temporary "
                         "sglang_legacy bypass and cannot declare canonical "
-                        "capabilities before the sglang_native PR-5 binder"
+                        "capabilities while sglang_native has no executable binder"
                     )
                 legacy_bypass.add(binding_id)
                 continue

sieval/cli/validation.py

@@ -308,7 +308,7 @@ def _validate_capabilities(cfg: dict, result: ValidationResult) -> None:
                 result.errors.append(
                     f"Model '{name}': the temporary sglang_legacy bypass "
                     "cannot declare canonical capabilities before the "
-                    "sglang_native PR-5 binder"
+                    "sglang_native executable binder is available"
                 )
             continue
 

sieval/core/models/dialect_registry.py

@@ -546,12 +546,12 @@ def bind_dialect(
 def _not_implemented_message(dialect_id: str) -> str:
     if dialect_id == "sglang_native":
         return (
-            "dialect 'sglang_native' has no PR-1 binder; the one-cycle SGLang "
-            "compatibility path remains an explicit legacy bypass"
+            "dialect 'sglang_native' has no executable binder; use the explicit "
+            "'sglang_legacy' compatibility bypass"
         )
     if dialect_id == "vllm_native":
-        return "dialect 'vllm_native' is reserved and explicitly deferred after #25"
-    return f"dialect {dialect_id!r} is reserved for a later #25 adapter PR"
+        return "dialect 'vllm_native' is reserved and has no executable binder"
+    return f"dialect {dialect_id!r} is reserved and has no executable binder"
 
 
 def dialect_registry_to_json() -> dict[str, JSONValue]:

sieval/core/models/dialects/anthropic_messages.py

@@ -205,7 +205,7 @@ def _validate_multimodal_config(options: object) -> None:
                 )
             ),
             "hosted_tools": Unsupported(
-                "PR 3 does not bind Anthropic server-hosted tools"
+                "anthropic_messages does not support Anthropic server-hosted tools"
             ),
             "structured_output": Supported(
                 DialectCapabilityBinding(
@@ -2167,7 +2167,10 @@ def validate_request(
                     "or 'none'",
                 )
             elif path == "tools.hosted":
-                audit.rejected(path, "Anthropic hosted tools are not active in PR 3")
+                audit.rejected(
+                    path,
+                    "Anthropic hosted tools are not supported by anthropic_messages",
+                )
             elif (
                 path == "structured_output.format"
                 and req.structured_output.format != "json_schema"
@@ -2203,7 +2206,8 @@ def validate_request(
                 else:
                     audit.rejected(
                         path,
-                        "PR 3 exposes no raw Anthropic request passthrough",
+                        "anthropic_messages does not support raw Anthropic request "
+                        "passthrough",
                     )
 
     def prepare(self, req: Request, audit: RequestAudit) -> PreparedRequest:

sieval/core/models/dialects/openai_chat.py

@@ -130,7 +130,7 @@ def _validate_multimodal_config(options: object) -> None:
             )
         ),
         "hosted_tools": Unsupported(
-            "PR 1 does not bind hosted tools on Chat Completions"
+            "openai_chat does not support hosted tools on Chat Completions"
         ),
         "structured_output": Supported(
             DialectCapabilityBinding(
@@ -766,7 +766,7 @@ def validate_request(
                     "openai_chat supports only summary='none' as a documented no-op",
                 )
             elif path == "tools.hosted":
-                audit.rejected(path, "hosted tools are not active in PR 1")
+                audit.rejected(path, "hosted tools are not supported by openai_chat")
             elif (
                 path == "input.modality.image.media_type"
                 and isinstance(req.input, ChatInput)

sieval/core/models/sglang_gen_model.py

@@ -232,14 +232,14 @@ async def arun(self, req: Request) -> Response:
     def with_dialect(self, dialect_id: str, runtime_plan: RuntimeBindingPlan) -> Model:
         del dialect_id, runtime_plan
         raise RuntimeError(
-            "sglang_legacy cannot rebind before the sglang_native PR-5 binder"
+            "sglang_legacy cannot rebind because sglang_native has no executable binder"
         )
 
     def with_provenance_plan(self, provenance_plan: RuntimeBindingPlan) -> Self:
         del provenance_plan
         raise RuntimeError(
-            "sglang_legacy has no runtime plan for provenance before the "
-            "sglang_native PR-5 binder"
+            "sglang_legacy has no runtime plan for provenance because "
+            "sglang_native has no executable binder"
         )
 
     def _legacy_lifecycle_owner(self) -> "SglangGenModel":

tests/unit/cli/test_validation.py

@@ -531,8 +531,8 @@ def test_sglang_legacy_bypass_rejects_capability_declarations(
     @pytest.mark.parametrize(
         ("dialect", "message"),
         [
-            ("sglang_native", "legacy bypass"),
-            ("vllm_native", "explicitly deferred"),
+            ("sglang_native", "no executable binder"),
+            ("vllm_native", "no executable binder"),
         ],
     )
     def test_registered_but_inactive_dialect_is_rejected(self, dialect, message):

tests/unit/core/models/dialects/test_anthropic_messages.py

@@ -194,6 +194,11 @@ def test_decision_row_is_complete_and_logprobs_are_explicitly_absent(self) -> No
         assert isinstance(CAPABILITY_DECISIONS["sampled_logprobs"], Unsupported)
         assert isinstance(CAPABILITY_DECISIONS["top_logprobs"], Unsupported)
         assert "does not return" in CAPABILITY_DECISIONS["sampled_logprobs"].reason
+        hosted = CAPABILITY_DECISIONS["hosted_tools"]
+        assert isinstance(hosted, Unsupported)
+        assert hosted.reason == (
+            "anthropic_messages does not support Anthropic server-hosted tools"
+        )
 
     def test_supported_capabilities_are_first_class_registry_outcomes(self) -> None:
         spec = DIALECT_SPECS["anthropic_messages"]
@@ -1379,7 +1384,7 @@ class TestPreIOGuards:
             ),
             (
                 _request(tools=ToolParams(hosted=(HostedToolSpec("web_search"),))),
-                "unavailable capability|not active",
+                "unavailable capability|not supported by anthropic_messages",
             ),
             (
                 _request(
@@ -1430,7 +1435,7 @@ class TestPreIOGuards:
                         "anthropic_messages", {"service_tier": "auto"}
                     )
                 ),
-                "no raw Anthropic request passthrough",
+                "does not support raw Anthropic request passthrough",
             ),
             (
                 _request(

tests/unit/core/models/dialects/test_openai_chat.py

@@ -13,7 +13,12 @@
 import pytest
 
 import sieval.core.models.dialects.openai_chat as openai_chat_module
-from sieval.core.models.capabilities import Capability, ReasoningOptions, Supported
+from sieval.core.models.capabilities import (
+    Capability,
+    ReasoningOptions,
+    Supported,
+    Unsupported,
+)
 from sieval.core.models.dialect import (
     DialectError,
     OutputContractError,
@@ -489,6 +494,14 @@ def test_legacy_capability_view_matches_supported_reasoning_controls(self) -> No
         assert Capability.Reasoning in dialect.capabilities
         assert Capability.ReasoningEffort in dialect.capabilities
 
+    def test_hosted_tool_rejection_uses_stable_dialect_wording(self) -> None:
+        decision = CAPABILITY_DECISIONS["hosted_tools"]
+
+        assert isinstance(decision, Unsupported)
+        assert decision.reason == (
+            "openai_chat does not support hosted tools on Chat Completions"
+        )
+
     def test_reasoning_config_rejects_non_noop_summary(self) -> None:
         decision = CAPABILITY_DECISIONS["reasoning"]
         assert isinstance(decision, Supported)

tests/unit/core/models/test_dialect_registry.py

@@ -264,16 +264,16 @@ def test_capability_decisions_are_available_only_for_active_dialects(self) -> No
         assert set(capability_decisions_for("anthropic_messages")) == set(
             CAPABILITY_KEYS
         )
-        with pytest.raises(DialectNotImplemented, match="later #25 adapter"):
+        with pytest.raises(DialectNotImplemented, match="no executable binder"):
             capability_decisions_for("google_genai")
 
 
 class TestDialectBinders:
     @pytest.mark.parametrize(
         ("dialect_id", "message"),
         [
-            ("vllm_native", "explicitly deferred"),
-            ("sglang_native", "legacy bypass"),
+            ("vllm_native", "no executable binder"),
+            ("sglang_native", "sglang_legacy.*bypass"),
         ],
     )
     def test_reserved_dialects_fail_with_named_error(