[Gastown] PR 20: Stress Testing

[jrf0110]

Parent: #204 | Phase 4: Hardening

Revised: Focus on container resource limits and multi-agent concurrency within a single container, rather than cloud-agent-next session throughput.

Goal

Validate the system under load and identify container resource limits.

Test Scenarios

• Simulate 30 concurrent polecats across 5 rigs in a single Town Container
• Measure DO→container latency under load (tool call round-trip time)
• Measure container resource usage (CPU, memory, disk) with N concurrent Kilo CLI processes
• Identify when to scale to larger instance types or shard to multiple containers
• Identify DO SQLite size limits and implement archival (closed beads → purge from DO)
• Test container crash/restart/restore cycles (ephemeral disk recovery)
• Measure review queue throughput under concurrent PR submissions

Metrics to Capture

• Tool call round-trip latency (p50, p95, p99)
• DO alarm scheduling accuracy under load
• SQLite query performance vs row count
• Container CPU/memory per Kilo CLI process
• Git clone/worktree creation time
• Container cold start time (fresh start vs restore from R2 snapshot)
• Memory/CPU usage per DO

Acceptance Criteria

• Load test harness for simulating concurrent polecats in a container
• Latency benchmarks documented
• Container resource usage profiled (CPU, memory, disk per agent)
• SQLite archival strategy implemented and tested
• Container crash/restore resilience validated
• Scaling thresholds documented (max agents per container by instance type)
• Performance bottlenecks identified and documented