Disables tagging elements in screenshots sent to vision models with their
accessibility refs. By default the navigator tags interactive elements and the
llm_text/llm_json formatter tags the elements in its prompt, each with a
small colored box reading its ref (e.g. @e7) so the model can correlate the
image with the accessibility-tree text. Tags are drawn only on the copy sent to
the model; rendered outputs and navigator video are unaffected.