docs: the resume line led with a p99 that was stale and mislabelled

[DanielWLiu07]

PLAN.md's resume line opened with "p99 37 us ingest-to-signal on recorded live sessions." That number was wrong in two independent ways.

1. It had drifted

The same committed captures measure 74–78 µs today. docs/bench/latency.md was updated to 77.5; this line and soak.md's cross-reference were not.

2. It's a service time, and the repo's headline says so

latency.md's entire finding is that the service p99 understates what a consumer waits by 13.3x. A resume line leading with an unqualified service p99 contradicts the repo's own strongest result.

The finding now replaces the figure:

Proved its own published p99 was a service time and 13.3x optimistic by replaying at the venue's real arrival schedule, and gated the invariant in CI.

That's the better line anyway. Anyone can report a percentile; almost nobody reports catching their own benchmark flattering itself.

Also

• 2.19M → ~2M messages/sec. 2.19 was a best-of-five; today's five runs span 1.80–1.90 on a busier machine. The ratio against the venue's 269 carries the point, not the third digit.
• Added the message counts (204,864 and 642,919) behind the z-scores, since a z with no n invites the question.
• soak.md's stale cross-reference fixed, with a note that the agreement between the two captures is the durable claim and the absolute figures move with the machine.