[zzylol]

Coordinator-side counterpart to ASAPCollector PR #532 (split/pr-retire-discipline-b-alerting).

Discipline B's global-threshold alerting (Monitor::rebroadcast's CMY slack-countdown → Action::Alert → the control-plane violation sink) retires per ASAPCollector/docs/design-gos-unified-edge-telemetry.md §11: "Alerting and any other query-time decision is made entirely at the backend against [the reconstructed sketch state]; the edge no longer makes alerting decisions itself." A collector/edge must never be the thing that fires an alert.

The whole-sketch ε-floor sampling law (p_i = 1/(1+ε²·rate_i)) was always a pure per-edge function with no cross-edge coordination or global-estimate dependency, so decoupling it from the retired countdown is a genuine simplification, not a workaround:

• Monitor drops known_value/departed_mass/alerted/tau/slack/gap entirely — just tracks each edge's last-reported rate.
• on_report answers ONLY the reporting edge with a freshly computed grant (no more re-broadcast to every registered edge on every report — that fan-out existed only because slack sizing depended on the full edge set).
• The retired "<2 edges ⇒ p=1" special case (needed only for the alert-fire safety proof under sampling noise) is gone: a lone high-rate edge is now sampled exactly like it would be in a larger edge set.

Removes alert.rs (AlertSink/global_threshold_violation) and the now-dead ViolationKind::GlobalThresholdCrossed. MonitorConfig.tau / Grant.LocalSlack stay on the wire (vestigial) for compatibility with the Go edge side, which still sends them.

Companion PR: ASAPCollector #532 — the edge side. Engine.Observe now reports its rate on its own periodic cadence (ReportEveryN observations), decoupled from any value/slack threshold.

Verified with a real cross-language run (Go e2edriver ↔ Rust monitor_coordinator_harness over live gRPC): the coordinator answers the edge's periodic rate report with a genuine computed sample_p grant.

Test plan

• data_plane::monitor::coordinator unit tests rewritten for the simplified per-edge model (registration grants nothing, report grants only the reporting edge, single-edge sampling, epoch reset, etc.)
• data_plane/tests/monitor_grpc.rs rewritten: dropped the two alert-firing tests, kept/added real gRPC round-trip tests for single-edge sampling, differentiated sample_p across edges, and report-scoped-to-reporter.
• Repurposed monitor_coordinator_harness to poll for a real coordinated-sampling grant instead of waiting for a retired alert; updated deploy/mvp-multinode/scripts/monitor_e2e.sh accordingly.
• Real cross-language e2e verified manually against the matching ASAPCollector worktree: MONITOR_GRANT edge=e2e-edge agg_id=1 sample_p=0.9975....
• cargo build/cargo test -p data_plane clean. cargo test -p control_plane has one pre-existing unrelated failure (Duration not in scope in optimizer/cost/mod.rs) present on the base branch before this change — not touched here.

🤖 Generated with Claude Code