The Served path emits no dispatch_event, so a serve cannot be attributed to a dispatcher generation.

Measured during #272 golden r2 triage (2026-09-08, read-only, at the run's sha): host B's entire
log had `grep -c DISPATCH_EV` = 0 while two serve arms demonstrably ran. Every question of the form
"did ONE generation serve this twice, or did TWO generations each serve once?" is therefore answered
by timing inference rather than by evidence — and my own phase-offset inference on that data was
REFUTED by its own falsifier (a3 inter-serve gaps 0-21 ms, median 4, scattered; no phase structure).

Ask: emit a dispatch_event on the Served path — at minimum for the Rest arm — carrying the serving
dispatcher's GENERATION. The claim path already does `served_w.fetch_add(1)`; the generation is
present at the point of service and is simply not written down.

Why it matters beyond tidiness: the sibling lane (two-dispatcher overlap) needs a DISTINCT-GENERATION
COUNT on the Served path to be decidable at all. Without this event that lane can only ever produce
timing arguments, which is precisely the class of evidence that already failed once here.

Scope: instrumentation only. No behaviour change to claiming, serving, or release. #272 did not
introduce or widen the gap — the claim path's only delta was the counter and the Rest arm's only
delta was a comment word — so this is not a regression fix, it is a missing instrument.

Type: ADDITION.