perf: scenario-based load generator using EDG SDK with metrics observability

[e7217]

Context

The NATS-level baseline (separate issue) measures the transport floor, but does not exercise EDG's actual ingest path: SDK serialization → core handler (internal/core/handler.go) → validation → platform.data.validated publish → JetStream → Telegraf → VictoriaMetrics. To answer realistic capacity questions ("can EDG sustain 10K msg/s with 100 simulated adapters?", "what's p99 end-to-end at burst?"), we need a scenario-driven load generator that uses the same SDK adapters do.

This complements ADR 0001 (#65) chaos tests: chaos validates correctness under failure, this validates performance under load.

Goals

1. Load generator binary at cmd/loadgen/:
   • Uses adapters/go/sdk so it exercises the same publish path as a real adapter
   • Configurable: number of simulated adapters, msg/s per adapter, payload size, ramp-up duration, total run time
   • Supports steady-state, ramp, and burst profiles
2. Observability:
   • Generator exports Prometheus metrics: sent count, send error rate, in-flight, p50/p95/p99 publish latency
   • Reuses/extends EDG core metrics for backlog, validation failures, JetStream pending
   • Grafana dashboard JSON committed under deploy/grafana/loadgen.json
3. Documentation at docs/perf/scenarios.md:
   • How to run, how to interpret metrics, how to capture a result snapshot
   • At least one recorded scenario result (e.g., "100 adapters × 100 msg/s, 5 min steady-state")
4. Reusable as a regression check — easy to rerun after changes to handler.go or stream config

Out of Scope

• Distributed load generation across multiple machines
• Concurrent chaos events (NATS restart, disk pressure) — those belong to reliability: Specify JetStream/ack semantics and add chaos tests #65's chaos tests
• Automated pass/fail thresholds in CI (manual analysis for now)
• Telegraf/VM-side tuning

Acceptance Criteria

• cmd/loadgen/ builds and runs against a local EDG stack
• At least three scenario profiles available via flags or config (steady, ramp, burst)
• Prometheus /metrics endpoint exposes generator-side latency histograms
• Grafana dashboard renders generator metrics alongside core EDG metrics
• docs/perf/scenarios.md includes one captured run with environment metadata and observations
• Generator uses adapters/go/sdk (no bypassing the SDK)

Open Questions

• Should payload generation include realistic per-asset distributions (mix of number/text/bool values), or fixed-shape payloads first? — fixed first, realistic as follow-up.
• Where to record historical results? — start with markdown snapshots, revisit if regressions become a regular concern.