Chaos CI Phase 1: pause, crash/recovery, disk-full

[ayushmi]

This PR introduces a self-contained chaos testing harness and CI workflow for Phase‑1 (single‑node) durability/availability validation.

What’s included

• Harness under AgentStateTesting/chaos with:
  • Workload: concurrent writers/readers driving public HTTP API.
  • Watcher: SSE consumer validating monotonic commit sequence and last-value agreement across nemeses.
  • Nemeses: pause/unpause (stall surrogate), crash/recovery, and disk‑full (tmpfs).
• Compose overlay: size‑limited tmpfs at /data, stable container name, test‑only Debian runtime with tools.
• CI: .github/workflows/chaos-ci.yml runs on PRs/nightly; prints container logs on failure.
• Metric assertions: ensure watch_drops_total{reason="overflow"} == 0, watch_clients{proto="sse"} >= 1, and watch_events_total > 0.

Notes

• No server/runtime code changes.
• SSE resumes aren’t counted yet (only gRPC does); left as a follow‑up if desired.
• Disk‑full is induced by filling tmpfs; we then free space and verify write recovery.

Follow‑ups (separate PRs)

• Add watch-stream soaks to nightly.
• Optional: track SSE resumes metric; clarify delete labeling in watch_events_total.
• Expand to true network partitions once clustering lands.

Signed-off-by: Ayush Mittal ayushsmittal@gmail.com