fix(embedding): make warmup best-effort so a keyless config doesn't fail startup

Eager warmup at startup constructed the embedding client unconditionally, so a deployment without an embedding API key crashed at boot instead of failing lazily on first embedding use as before. Keep failing fast on a bad tokenizer spec (ValidationException), but log and defer any other construction error.

src/embedding_client.py

@@ -805,9 +805,23 @@ class EmbeddingClient:
         ``hf:`` tokenizer specs fetch from the Hugging Face Hub; doing that on
         the first request would stall the event loop under the singleton lock.
         Calling this from the API and deriver lifespans moves any tokenizer
-        download off the request path and fails the process fast on a bad spec.
+        download off the request path and fails the process fast on a bad
+        tokenizer spec. Any other construction error (e.g. a missing API key)
+        is logged and deferred to first use, so deployments that never embed
+        keep starting up exactly as they did before warmup existed.
         """
-        self._get_client()
+        try:
+            self._get_client()
+        except ValidationException:
+            # An explicitly configured tokenizer spec that cannot load is a
+            # misconfiguration of this feature; keep failing fast.
+            raise
+        except Exception:
+            logger.warning(
+                "Embedding client warmup failed; embedding calls will fail "
+                + "until the embedding configuration is fixed",
+                exc_info=True,
+            )
 
     def _get_settings_signature(self) -> tuple[object, ...]:
         runtime_config = self._resolve_runtime_config()

tests/llm/test_embedding_client.py

@@ -1282,3 +1282,56 @@ def test_settings_signature_tracks_tokenizer(
     after = wrapper._get_settings_signature()  # pyright: ignore[reportPrivateUsage]
 
     assert before != after
+
+
+def test_warmup_defers_missing_api_key_instead_of_failing_startup(
+    monkeypatch: pytest.MonkeyPatch,
+    caplog: pytest.LogCaptureFixture,
+) -> None:
+    """A keyless embedding config must not crash startup: warmup logs and
+    defers the failure to first use, matching pre-warmup lazy behavior."""
+    from src.embedding_client import EmbeddingClient
+
+    wrapper = EmbeddingClient()
+    monkeypatch.setattr(EmbeddingClient, "_instance", None)
+    monkeypatch.setattr(EmbeddingClient, "_instance_signature", None)
+    monkeypatch.setattr(
+        wrapper,
+        "_resolve_runtime_config",
+        lambda: EmbeddingModelConfig(
+            transport="openai",
+            model="text-embedding-3-small",
+            api_key=None,
+        ),
+    )
+
+    with caplog.at_level("WARNING", logger="src.embedding_client"):
+        wrapper.warmup()
+
+    assert EmbeddingClient._instance is None  # pyright: ignore[reportPrivateUsage]
+    assert "warmup failed" in caplog.text
+
+
+def test_warmup_still_fails_fast_on_bad_tokenizer_spec(
+    monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """An explicitly configured tokenizer spec that cannot load is a
+    misconfiguration of the tokenizer feature and must keep failing startup."""
+    from src.embedding_client import EmbeddingClient
+
+    wrapper = EmbeddingClient()
+    monkeypatch.setattr(EmbeddingClient, "_instance", None)
+    monkeypatch.setattr(EmbeddingClient, "_instance_signature", None)
+    monkeypatch.setattr(
+        wrapper,
+        "_resolve_runtime_config",
+        lambda: EmbeddingModelConfig(
+            transport="openai",
+            model="text-embedding-3-small",
+            api_key="test-key",
+            tokenizer="tiktoken:",
+        ),
+    )
+
+    with pytest.raises(ValidationException, match="requires an encoding name"):
+        wrapper.warmup()