Export queue drop and drain metrics for replay and overload analysis

[jillesvangurp]

Summary

The current OTLP instrumentation shows queue depths and ingest rates, but it does not expose the overload signals we actually needed during the OpenSky replay benchmark.

During the 2026-04-08 benchmark run:

• the 10 Hz replay expanded a five-minute Germany capture to 1,581,177 replayed updates
• the runtime started dropping work on the decision queue long before host CPU was saturated
• queue drops were visible in logs but not as first-class metrics, which made it harder to dashboard overload behavior directly in SigNoz

Missing telemetry

Add explicit OTLP metrics for:

• native queue dropped item counter
• decision queue dropped item counter
• event-bus outbound dropped item counter if applicable
• websocket outbound dropped item counter if applicable
• queue enqueue/dequeue counters by stage
• queue drain rate by stage
• stage worker concurrency / active worker count where measurable
• collision evaluation count and fence evaluation count as separate counters from the generic processing timings
• per-stage success, drop, and skip outcomes so dashboards can show where overload is happening without reading logs

Acceptance criteria

• queue drops are exported as monotonic counters with low-cardinality stage labels
• default SigNoz dashboards include overload panels for drop rate and cumulative drop count
• the local benchmark workflow can assert that selected runs remain drop-free without parsing container logs
• docs are updated with the new metrics and dashboard intent

Context

This came out of the OpenSky Germany replay benchmark that now lives in connectors/replay/benchmarks/opensky-germany-2026-04-08/.

[jillesvangurp]

PR #44 adds OTEL export for native and decision queue drop counters plus tests and docs. It advances the overload-telemetry slice from this issue, but it does not fully satisfy the remaining acceptance criteria around the broader drop/drain metric set, dashboards, and benchmark assertions without log parsing, so I left this issue open.