[gvonnessi]

Summary

Follow-up to #20. The test that shipped with the exemption did not pin it.

Peer and probe both ran on a 60s step, so the first probe to land set hist.last() to its own timestamp and the next was admitted by the ordinary throttle on the boundary — last.t > now - INTERVAL is false at equality. The session sustained its own cadence, the peer never regained the slot, and from the second write onward the exemption was not what kept the probes landing.

This was found by mutation, not inspection: an exemption that fired twice and then did nothing passed the old test unchanged. Only outright removal failed it — the regression least likely to occur, and the one review would catch anyway.

• The peer now writes one second before each probe on a 61s step. At every probe the session's own last sample is 61s old (opening the per-session gate) while the shared slot was claimed a second ago (so the ordinary throttle would refuse it). Each probe lands only on the exemption.
• A final uncontended write shows the hold releasing: with the switch proven it wins the ordinary slot unaided and is recorded as spend rather than demoted to a probe.
• Narrows the gate from "the counter differs" to "the counter climbed", so it agrees with hold, which releases on u > first_move. A slot spent on a counter that went down buys evidence the release can never use. Unreachable with real cumulative counters; it stops the two rules disagreeing.

Test plan

• cargo fmt --check, cargo clippy --all-targets -- -D warnings, cargo build, cargo test --all (85 passed)
• Partial break — exemption fires twice then stops: passed before, now fails (got 5 against the 40+ the schedule produces). This is the regression the old test could not see.
• Full removal — still fails (got 3)
• Re-confirmed both after the directional narrowing