The framing that 'speculation enabled' and 'LCP improved' are separate claims is the part worth underlining, because the demo click always looks instant while the entry page quietly pays for it. The failure-mode list is practical too — especially action URLs such as add-to-cart and checkout slipping into the candidate set, and several stacked speculation sources fighting over the same mobile radio. The two-clock measurement plan, cold-entry LCP against next-navigation feel, seems like the right discipline. Does the scheduled lab run after a settings change usually catch regressions that a one-off PSI check misses, or is field data still the deciding signal?