[Bug] Some common uses of ActiveModel fail in workflows due to sync construct use

首先，在 workflow 中复现所示的 ActiveModel::Model 和 ActiveModel::Attributes 示例，并检查 trace 中的 attribute_methods.rb 和 illegal_call_tracer.rb 帧。确定哪些 ActiveModel 操作会调用 ConcurrentMap 或 Thread::Mutex，然后通过回归覆盖定义并验证对 workflow 安全的行为；当常见的 ActiveModel 使用不再触发 workflow 非确定性错误时，即表示完成。

Describe the bug

Many common uses of ActiveModel use ConcurrentMap under the hood which uses Thread::Mutex which is forbidden in workflows.

We believe this can be replicated simply by accessing an attribute that doesn't exist. We also believe this can be replicated using a model set like:

module MyCompany
  module Messages
    class Foo
      include Message

      attribute :some_field_1, :string
      attribute :some_field_2, :string
      attribute :some_field_3, ObjectType.for(Array), :default => []
      attribute :some_field_4, :boolean, :default => false
      attribute :some_field_5, :boolean, :default => false
    end
  end

  module Message
    extend ActiveSupport::Concern
    include ActiveModel::Model
    include ActiveModel::Attributes
    include ActiveModel::Serializers::JSON

    included do
      def as_json(options = {})
        super(options).merge(::JSON.create_id => self.class.name)
      end
    end

    class_methods do
      def json_create(data)
        new(**data.except(JSON.create_id))
      end
    end
  end
end

This can give stack traces like:

W, [2025-10-24T12:22:16.291089 #70269]  WARN -- : Cannot access Thread::Mutex synchronize from inside a workflow, reason: disallowed. If this is known to be safe, the code can be run in a Temporalio::Workflow::Unsafe.illegal_call_tracing_disabled block. {:attempt=>1, :namespace=>"default", :run_id=>"019a1773-c113-74ff-924d-7c51d86cd682", :task_queue=>"deliverable-orchestration", :workflow_id=>"59824eb54dc4fa489c15c372", :workflow_type=>"OrchestrationWorkflowV2"} (Temporalio::Workflow::NondeterminismError)
/path/to/gems/temporalio-0.6.0-arm64-darwin/lib/temporalio/internal/worker/workflow_instance/illegal_call_tracer.rb:112:in `block in initialize'
/path/to/gems/concurrent-ruby-1.3.4/lib/concurrent-ruby/concurrent/collection/map/mri_map_backend.rb:25:in `compute_if_absent'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_methods.rb:400:in `attribute_method_matchers_matching'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_methods.rb:506:in `matched_attribute_method'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_methods.rb:494:in `respond_to?'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_assignment.rb:48:in `_assign_attribute'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_assignment.rb:42:in `block in _assign_attributes'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_assignment.rb:41:in `each'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_assignment.rb:41:in `_assign_attributes'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attribute_assignment.rb:34:in `assign_attributes'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/model.rb:81:in `initialize'
/path/to/gems/activemodel-6.1.7.10/lib/active_model/attributes.rb:77:in `initialize'
/path/to/models/my_models.rb:99:in `new'

And for cases where it's an invalid attribute access via method_missing, it can give traces like:

/path/to/gems/temporalio-1.0.0-arm64-darwin/lib/temporalio/internal/worker/workflow_instance/illegal_call_tracer.rb:112:in `block in initialize'
/path/to/gems/concurrent-ruby-1.3.4/lib/concurrent-ruby/concurrent/collection/map/mri_map_backend.rb:25:in `compute_if_absent'
/path/to/gems/activemodel-6.1.7.3/lib/active_model/attribute_methods.rb:400:in `attribute_method_matchers_matching'
/path/to/gems/activemodel-6.1.7.3/lib/active_model/attribute_methods.rb:506:in `matched_attribute_method'
/path/to/gems/activemodel-6.1.7.3/lib/active_model/attribute_methods.rb:468:in `method_missing'
my_workflow_class.rb:52:in `block in execute'
my_workflow_class.rb:39:in `loop'

ConcurrentMap use is so pervasive in ActiveModel, it is unreasonable for us to ask users not to use anything relying on it. Also, it may be unreasonable to just alter our illegal trace detector to check backtrace for the active model because the use of mutexes can cause an issue like we hit with loggers (see https://temporal.io/blog/temporal-ruby-crash-proof-fibers#implicitly-used-sync-constructs).

The best solution may be a WorkflowSafeObject type of mixin that surrounds every call with Workflow::Unsafe::durable_scheduler_disabled.