We need performance test cases which measure the effectiveness of speculative parsing. See also Bug 92224.
I'm not sure how we'd do this?
I think this is also WONTFIX for now. Ideally page load test benchmarks cover this.