Changelog

Re-architected the Playfair test for full reproducibility. The 9-phase bash script is now a thin wrapper around a Terraform module that declares every state transition (cluster, image builds, image imports, latency injection, manifests, orchestrator, results) as a tracked resource with content-hashed triggers.

• Terraform module — tests/playfair/terraform/ manages 25 resources via null_resource + local-exec. Idempotent re-runs only re-execute the resources whose source-tree SHA-256 changed; terraform destroy tears down the entire cluster + tc netem state cleanly.
• GitHub Actions workflow — .github/workflows/playfair.yml: nightly @ 03:00 UTC, on push to main, and workflow_dispatch with epoch + latency overrides. Builds all images on the runner, applies Terraform against a local k3d cluster, uploads the report as an artifact, and (on main) commits the rendered HTML back so it's published via Pages.
• Documentation page — new /playfair.html with full what / why / when / produces / represents / means / how conditions vary walkthrough, plus tuning + debugging table.
• Polished report — generate-playfair-report.js rewritten with inline SVG charts (stacked area timeline of perceives/stores/routes/syncs, per-agent activity sparklines, per-region token-usage bars), a "runtime config" panel with commit/runner/branch metadata, and an embedded "how to read" explainer.
• Bug fixes baked in: (1) Prisma engine resolution failure on Alpine arm64 — PRISMA_QUERY_ENGINE_LIBRARY hard-pinned in 03-services.yaml; (2) thalamus-router crash from sharing one NATS JetStream consumer across 3 regions — consumer names now scoped by ECCA_REGION; (3) k3d image import deadlock under concurrent Terraform workers — imports serialised in a single shell loop; (4) empty playfair-results.json when the orchestrator pod terminated before kubectl cp — fallback extracts JSON between the ═══ RESULTS JSON ═══ marker and the next banner so the trailing summary doesn't corrupt parsing; (5) Node 24 ESM/CJS clash — tests/playfair/package.json now declares "type": "commonjs" overriding the parent tests/package.json so the host-side report generator runs.
• Run modes: ./tests/playfair/run.sh (full), --skip-images (reuse cached), --epochs N, --skip-latency, --destroy, --plan.
• Verified runs (against 618eac8 on macOS aarch64 / Docker Desktop / k3d v5.8.3):
  • Run A — 2026-05-09T13:36:25Z → 13:37:17Z (orchestrator 56s, terraform apply ≈ 8 min from-scratch): 50 epochs, 6 agents, 3 regions, 152 perceptions / 43 stores / 31 routes, 1 residue (resolved), 31 cross-region migrations, 9 scripted scenarios, verdict FAIR on every epoch.
  • Run B — 2026-05-09T13:47:13Z → 13:48:05Z (orchestrator 51.7s, full re-run with rebuilt orchestrator image): 50 epochs, 152 perceptions / 49 stores / 33 routes / 24 syncs, 1 residue (resolved), 33 migrations, verdict FAIR on every epoch (unfairEpochs: []).
  • Both runs produced a 60 KB self-contained HTML report (8 inline SVG charts) and a 113 KB results JSON; terraform destroy -auto-approve tore the cluster down cleanly between runs.

See the latest run at /playfair-report.html and the docs at /playfair.html.

Complete Kubernetes-based test harness that runs 6 agents across 3 regions with asymmetric resource profiles, proving the variable-bitrate thesis from the dev blog in actual multi-node infrastructure.

• 3-region K8s cluster — k3d with labeled agent nodes: region-storage (cheap storage, throttled compute), region-compute (cheap compute, throttled storage), region-bandwidth (cheap bandwidth, both others throttled). Each region deploys its own medulla-pow, hippocampus-dag, and cortex-evm with tuned resource limits and ECCA_DIFFICULTY settings
• Shared infrastructure — Postgres 16, Redis 7, NATS 2.10 (JetStream), MinIO in a shared namespace; all regions reference via cross-namespace DNS
• 6 specialized agents — Archivist-Alpha/Beta (memory/human sleeves in storage), Inference-Prime/Echo (AI sleeves in compute), Router-Nexus/Sentinel (mining/memory sleeves in bandwidth). Each has unique CPV coefficients, tick rates, and behavioral profiles
• 9 scripted scenario events — spot-preemption at epoch 5, respawn in different region at epoch 8, needlecast back at epoch 15, drift-spike/sync-recovery at epochs 20/25, shard-loss residue injection at epoch 30, cross-region needlecast at epoch 35, epoch-surge stress test at epoch 40, return migration at epoch 45
• Orchestrator — orchestrator.js: Node.js game engine running 50 epochs with per-agent perceive/store/route/sync actions at variable rates, fairness auditing every epoch, token usage tracking (compute/storage/bandwidth), and full results JSON output
• Report generator — generate-playfair-report.js produces playfair-report.html in the cyberpunk neon theme with region comparison cards, agent stat cards with token usage/drift/CPV, ASCII epoch timeline sparklines, needlecast flow log, scenario event timeline, and fairness verdict banner
• Master runner — 9-phase run.sh (cluster → build → shared infra → regions → contracts → services → game → results → cleanup) with --skip-cluster, --skip-build, --epochs N flags
• Docker images — Dockerfile.orchestrator (node:20-alpine), Dockerfile.service (generic TS service with SERVICE build arg)
• Contract deployer — K8s Job that waits for cortex-evm RPC availability, then runs npx hardhat run scripts/deploy.ts --network cortex

Run: pnpm test:playfair or ./tests/playfair/run.sh --epochs 20

• Resolved all build errors across TypeScript packages, services, workers, and Go forks
• Fixed Prisma schema field mismatches (Anchor, Residue, Epoch, AuditLog models)
• Fixed @ecca/proto export mismatches — added standalone re-exports for EPOCH_INTERVAL_MS, DRIFT_MAX_DEFAULT, etc.
• Fixed @ecca/crypto type mismatches — hex encoding for coherenceRoot() callers
• Monorepo structure: pnpm workspace, turborepo, Docker Compose (24 services), Helm chart stubs