fix(hf-space): use provider="auto" for InferenceClient (new routing layer)

User confirmed:
- All HF Inference Providers enabled (incl. Featherless)
- Access granted to gated google/gemma-2-9b-it
- Featherless AI shown as the provider on the model's playground

Yet the Space call still returns "model_not_supported". Root cause:
the legacy `InferenceClient(model=...)` constructor without provider=
defaults to hf-inference-only routing, which doesn't serve most newer
models even when the user has access through partners.

Modern pattern (HF Inference Providers, intro'd late 2024) requires
`provider="auto"` to opt into automatic partner selection. With this,
the client picks Featherless for Gemma 2 9B because that's the live
provider for the model on the user's account.

Pure config change — no schema/contract changes — 31 tests still pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

app.py

@@ -302,7 +302,19 @@ def _call_huggingface(system_block: str, user_prompt: str) -> str:
             "https://huggingface.co/settings/tokens). Then restart the Space. "
             "Until then, pick a different model from the dropdown."
         )
-    client = InferenceClient(model=HF_MODEL_ID, token=token, timeout=120)
+    # `provider="auto"` opts into the modern HF Inference Providers
+    # routing layer (introduced 2024-Q4), which picks the right partner
+    # (featherless-ai / together-ai / hf-inference / etc.) for the model
+    # automatically. Without this flag, InferenceClient falls back to
+    # the legacy hf-inference-only path, which doesn't serve most newer
+    # models and returns a misleading "model not supported" error even
+    # when the user has all providers enabled and access to the model.
+    client = InferenceClient(
+        model=HF_MODEL_ID,
+        token=token,
+        provider="auto",
+        timeout=120,
+    )
     try:
         resp = client.chat_completion(
             messages=[