Measure critical-path and deferred latency separately

When blocking work moves behind a deferred endpoint, measure critical-path page latency and deferred-endpoint latency as separate contracts. A new endpoint may have no pre-release series and may remain slow while the visible page improves; treating a missing pre-period as zero or calling the endpoint itself faster is invalid.

A like-for-like delta still requires the same environment, workload class, and measurement surface. Report an architectural critical-path improvement separately from historical endpoint performance.