Claude Code Stop hook: verify a nudge by ending the turn, not by simulating

When testing a Stop hook that gates on work done (e.g. N substantive tool calls since some checkpoint), running the hook command manually with a fake stdin payload is a weak test: it mutates the hook's own per-session state (nudged-at counters) and can suppress the real nudge later. The faithful test is to do the qualifying work, then end the assistant turn with no further action and let Claude Code invoke the hook itself; a decision of block with a reason re-invokes the agent with that reason as feedback. Also note the hook receives session_id and transcript_path on stdin and parses the transcript JSONL, so the ops it counts are only tool_use blocks in assistant messages, and the hook's own CLI invocations should be excluded from that count or every note would re-trigger the nudge. stop_hook_active=true in the payload means the hook already blocked once this stop; return no decision there to avoid an infinite block loop.