ops: load testing and a capacity plan

[joelpeace48-cell]

Problem

Nobody knows what this service can handle. The pitch targets high-volume
inference gateways, so "how many attestations per second" is a question a pilot
partner will ask in the first conversation, and the honest answer today is that
it has never been measured.

What to do

• Load tests with k6 or Artillery against realistic mixes: heavy usage
  ingestion, moderate reads, occasional large exports.
• Find the actual bottleneck — it is most likely RPC rate limits or transaction
  throughput rather than the Node process, and knowing which changes what you
  optimise.
• Measure latency percentiles, not averages. The p99 is what a partner notices.
• Establish per-instance capacity and a scaling model, including what does
  not scale horizontally (the indexer is probably singleton).
• Re-run in CI on a schedule to catch regressions.

Acceptance criteria

• Load test suite with realistic scenarios
• Bottleneck identified and documented
• p50/p95/p99 latency recorded per endpoint
• Capacity and scaling model published, including singleton components
• Scheduled regression runs

Notes

Naming the components that cannot scale horizontally is the useful part. The
indexer almost certainly cannot, and discovering that during a traffic spike is
the wrong time.