FID measured the delay before the browser could start handling the first input. INP measures the round-trip latency of every interaction, then reports the worst (98th percentile).
The implication: a fast initial input does not save you. Long tasks during scroll, click handlers that block, or layout thrashing during interaction all count.
Tools: PerformanceObserver with event entry type captures real interactions; the Web Vitals JS library reports INP for the page.