defer caching nnf ops to avoid race

Author: ledwards2225

barretenberg/cpp/src/barretenberg/stdlib_circuit_builders/ultra_circuit_builder.cpp

@@ -1576,15 +1576,20 @@ std::array<uint32_t, 2> UltraCircuitBuilder_<ExecutionTrace>::queue_partial_non_
     const uint32_t hi_0_idx = this->add_variable(hi_0);
     const uint32_t hi_1_idx = this->add_variable(hi_1);
 
-    // Add witnesses into the multiplication cache (duplicates removed during circuit finalization)
+    // Add witnesses into the multiplication cache (duplicates removed during circuit finalization).
+    // In cursor mode, defer to per-thread buffer to avoid races on the shared vector.
     cached_partial_non_native_field_multiplication cache_entry{
         .a = input.a,
         .b = input.b,
         .lo_0 = lo_0_idx,
         .hi_0 = hi_0_idx,
         .hi_1 = hi_1_idx,
     };
-    cached_partial_non_native_field_multiplications.emplace_back(cache_entry);
+    if (this->get_variable_cursor() != this->VARIABLE_CURSOR_DISABLED) {
+        deferred_non_native_field_muls_[get_parallel_thread_index()].emplace_back(cache_entry);
+    } else {
+        cached_partial_non_native_field_multiplications.emplace_back(cache_entry);
+    }
     return std::array<uint32_t, 2>{ lo_0_idx, hi_1_idx };
 }
 

barretenberg/cpp/src/barretenberg/stdlib_circuit_builders/ultra_circuit_builder.hpp

@@ -218,13 +218,31 @@ class UltraCircuitBuilder_ : public CircuitBuilderBase<typename ExecutionTrace_:
     };
     std::vector<std::vector<DeferredRangeConstraint>> deferred_range_constraints_; // per-thread
 
+    // Deferred non-native field multiplications for parallel construction. In cursor mode,
+    // these are buffered per-thread to avoid races on the shared vector.
+    std::vector<std::vector<cached_partial_non_native_field_multiplication>> deferred_non_native_field_muls_;
+
     /**
      * @brief Initialize deferred buffers for N threads.
      */
     void init_deferred_buffers(size_t num_threads)
     {
         deferred_lookup_gates_.resize(num_threads);
         deferred_range_constraints_.resize(num_threads);
+        deferred_non_native_field_muls_.resize(num_threads);
+    }
+
+    /**
+     * @brief Replay all deferred non-native field multiplications into the shared cache.
+     */
+    void apply_deferred_non_native_field_muls()
+    {
+        for (auto& thread_buf : deferred_non_native_field_muls_) {
+            for (auto& entry : thread_buf) {
+                cached_partial_non_native_field_multiplications.emplace_back(entry);
+            }
+            thread_buf.clear();
+        }
     }
 
     /**
@@ -467,6 +485,7 @@ class UltraCircuitBuilder_ : public CircuitBuilderBase<typename ExecutionTrace_:
         // Replay deferred operations
         apply_deferred_lookup_gates();
         apply_deferred_range_constraints();
+        apply_deferred_non_native_field_muls();
         {
             size_t total_deferred = 0;
             for (size_t t = 0; t < this->deferred_assert_equals_.size(); t++) {