fix(so): validate required so_args keys before building the matcher

build() already rejects the unsupported serialized and bound_token_ids
keys with a clear message, but read the two required keys with a bare
subscript. A so_args missing schema raised KeyError('schema') on the
scheduling thread, and one missing schema_type raised inside the worker
closure, where try_result() swallowed it and reported the request as
"xgrammar unknown Error: 'schema_type'" — an API misuse surfacing as an
opaque grammar failure.

Check both keys up front, next to the existing so_args validation, and
read schema_type on the caller's thread like schema so the worker
closure no longer touches so_args.

deepinfer/structure_output/so_matcher_builder.py

@@ -61,6 +61,10 @@ class SOMatcherBuilder:
             raise ValueError("Offline structured output does not accept serialized grammars; provide schema_type and schema")
         if "bound_token_ids" in so_args:
             raise ValueError("bound_token_ids is no longer supported; use bound_token_id with a single non-negative integer")
+        missing = [key for key in ("schema_type", "schema") if key not in so_args]
+        if missing:
+            raise ValueError(f"so_args is missing required key(s) {missing}; provide schema_type and schema")
+        schema_type = so_args["schema_type"]
         schema = so_args["schema"]
 
         trigger_tokens = so_args.get("trigger_token_ids")
@@ -76,7 +80,7 @@ class SOMatcherBuilder:
             return XGrammarMatcherFuture(task=failed)
 
         def _build() -> XGrammarMatcher:
-            ctx = self._compiler.dispatch_and_compile(so_args["schema_type"], schema)
+            ctx = self._compiler.dispatch_and_compile(schema_type, schema)
             matcher = XGrammarMatcher.from_compiled(
                 ctx,
                 self._vocab_size,