test(translation): encode the baseten steady state and registration-coverage completeness (critic-wave1a M2, verifier-wave1a F3)

The drop canary gains a registration half: negative asserts that baseten
is absent from the dispatch Literal and all four pipeline tables, so naive
registration goes red with or without a corpus row (the verifier's variant
B experiment sailed through green before this). And a completeness gate
pins Literal == every pipeline table == dedicated-gate providers + the
compat_sdk SPECS corpus, with PROFILES == ALLOWED == SPECS, so ANY future
wave registering a provider without differential coverage fails here, not
just baseten.

tests/test_litellm/translation/test_differential_compat_sdk_request.py

@@ -13,6 +13,7 @@ the v1 configs at HEAD, over every model-map row where the provider has any.
 
 import copy
 import json
+from typing import get_args
 
 import pytest
 
@@ -20,6 +21,9 @@ import litellm
 from litellm.exceptions import UnsupportedParamsError
 
 from litellm.translation import translate_chat_request
+from litellm.translation.dispatch import Provider
+from litellm.translation.engine import pipeline
+from litellm.translation.providers import compat_sdk
 from litellm.translation.providers.compat_sdk import params as csp
 
 from ._compat_sdk_corpus import (
@@ -244,6 +248,36 @@ def _request_rows():
     )
 
 
+def test_registered_providers_have_differential_coverage() -> None:
+    """Green-but-untested registration is impossible (verifier-wave1a F3):
+    every dispatch Literal member is registered in every pipeline table,
+    the compat_sdk family registry equals the corpus SPECS exactly, and a
+    provider outside the family must be in the dedicated-gates set below.
+    Add a provider to that set ONLY in the commit that adds its
+    differential corpus, naming the gate files."""
+    providers = set(get_args(Provider))
+    assert providers == set(pipeline._SERIALIZERS)
+    assert providers == set(pipeline._RESPONSE_PARSERS)
+    assert providers == set(pipeline._RESPONSE_DIALECTS)
+    assert set(pipeline._RAW_GUARDS) <= providers
+    assert set(compat_sdk.PROFILES) == set(SPECS)
+    assert set(compat_sdk.ALLOWED) == set(SPECS)
+    dedicated_gates = {
+        "anthropic",  # test_differential_anthropic_{request,response,stream}
+        "bedrock_converse",  # test_differential_bedrock_*
+        "bedrock_invoke",  # test_differential_bedrock_*
+        "openai_compat",  # test_differential_openai_*
+        "vertex_ai",  # test_differential_google_*
+        "gemini",  # test_differential_google_*
+        "vertex_anthropic",  # test_differential_google_*
+        "azure",  # test_differential_azure_*
+        "azure_ai",  # test_differential_azure_ai_request + azure stream/response
+        "azure_ai_anthropic",  # test_differential_azure_ai_request (Claude route)
+        "xai",  # test_differential_xai_*
+    }
+    assert providers == dedicated_gates | set(SPECS)
+
+
 @pytest.mark.parametrize("provider,name", _request_rows())
 def test_v2_request_matches_v1(provider: str, name: str) -> None:
     case = corpus_for(provider)[name]

tests/test_litellm/translation/test_differential_compat_sdk_stream.py

@@ -16,6 +16,7 @@ the bottom pins the evidence so the drop reason stays true at HEAD.
 import copy
 import json
 import time
+from typing import get_args
 
 import pytest
 from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
@@ -23,6 +24,9 @@ from openai.types.chat.chat_completion_chunk import ChatCompletionChunk
 from litellm.litellm_core_utils.litellm_logging import Logging
 from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
 
+from litellm.translation.dispatch import Provider
+from litellm.translation.engine import pipeline
+
 from ._compat_sdk_corpus import PROVIDERS
 from .test_differential_openai_stream import MODEL, STREAMS, USAGE_STREAM, _v2_chunks
 
@@ -85,10 +89,20 @@ def test_baseten_drop_canary(frozen_ambient) -> None:
     dedicated legacy decoder (``handle_baseten_chunk``,
     streaming_handler.py:1246-1248) instead of the default openai arm, so
     its v1 stream behavior is not the openai dialect and cannot be honestly
-    registered. If this canary fails, the legacy branch is gone at HEAD —
+    registered. If the DRIFT half fails, the legacy branch is gone at HEAD —
     re-evaluate porting baseten (request/response sides are trivial:
-    own list, mct rename, user supported)."""
+    own list, mct rename, user supported). The REGISTRATION half makes the
+    steady state loud (critic-wave1a M2 / verifier F3): registering baseten
+    anywhere requires deliberately deleting these negative asserts, which
+    forces the registrar past the evidence above."""
+    # drift half: the legacy branch still exists and still diverges
     assert hasattr(CustomStreamWrapper, "handle_baseten_chunk")
     v1_baseten = _v1_chunks("baseten", STREAMS["text"])
     v1_openai = _v1_chunks("custom_openai", STREAMS["text"])
     assert _norm(v1_baseten) != _norm(v1_openai)
+    # registration half: naive registration fails HERE, corpus row or not
+    assert "baseten" not in get_args(Provider)
+    assert "baseten" not in pipeline._SERIALIZERS
+    assert "baseten" not in pipeline._RESPONSE_PARSERS
+    assert "baseten" not in pipeline._RESPONSE_DIALECTS
+    assert "baseten" not in pipeline._RAW_GUARDS