fix: fail closed when grammar parsing fails (#19051)

When response_format: json_schema is requested but grammar parsing
fails (e.g., invalid regex pattern), the server now returns HTTP 400
error instead of silently generating unconstrained output.

Previously, common_sampler_init() would return a valid sampler even
when grammar parsing failed (grmr=nullptr). This caused the server
to continue with unconstrained generation, violating the structured
output contract.

Now, if grammar was requested but parsing failed, common_sampler_init()
returns nullptr, triggering the existing error handling path.

Fixes #19051

common/sampling.cpp

@@ -191,6 +191,11 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
     if (params.grammar.compare(0, 11, "%llguidance") == 0) {
 #ifdef LLAMA_USE_LLGUIDANCE
         grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
+        // fail closed: if llguidance grammar parsing failed, return nullptr
+        if (grmr == nullptr) {
+            llama_sampler_free(chain);
+            return nullptr;
+        }
 #else
         GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
 #endif // LLAMA_USE_LLGUIDANCE
@@ -247,6 +252,12 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, st
              } else {
                  grmr = llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
              }
+             // fail closed: if grammar was requested but parsing failed, return nullptr
+             // this ensures the server returns an error instead of generating unconstrained output
+             if (grmr == nullptr) {
+                 llama_sampler_free(chain);
+                 return nullptr;
+             }
         }
     }