Part of Epic: Factory Verification & Environments (#141). Depends on the EnvironmentProvider (#142) and stack deployer (#143).

Why

"Fully tested end to end" includes behavior under load, not just a single happy-path request. The Factory needs to drive realistic traffic at a deployed stack and decide pass/fail against SLOs — otherwise a feature that works for one request but collapses at concurrency ships green.

What

1. Load runner — run a load generator (k6 recommended; Locust acceptable) against the environment's endpoints. Run it from a Cloudflare Container in the environment (or an external CI runner) so traffic exercises the real deployed Workers/Containers; Cloudflare Workers autoscale and Container instances scale within the env's cap.
2. Profile — a declarable load profile (per the verification-stack descriptor or a sidecar file): target endpoints + request shapes, RPS / virtual users / ramp / duration.
3. SLO gate — assert thresholds: p95/p99 latency, error rate, throughput floor. Return a structured pass/fail with the measured numbers.
4. Evidence — capture the run summary (latency histogram, error rate, RPS) as evidence JSON so the merge-gate and humans can inspect it.
5. Export a runLoad(env, profile): LoadResult surface from src/index.ts.

Acceptance

• Load run drives the provisioned environment's endpoints with a configurable profile.
• SLO gate returns pass/fail with measured p95/error-rate/throughput.
• Evidence JSON captured; thresholds are declarative.
• Exported from src/index.ts.

✅ End-to-end verification — Factory success criteria (REQUIRED)

"Compiles" is not done. Fire real load at a real deployed service and prove the SLO gate flips correctly.

Run it

npm run build && npm test

CLOUDFLARE_API_TOKEN=... npx tsx test/e2e/load-harness.e2e.ts

The E2E must:

1. Provision (EnvironmentProvider port + Cloudflare ephemeral-environment provider (Workers for Platforms + Containers); K8s/Daytona optional #142) + deploy (Verification-stack descriptor contract + deployer: repos declare their full stack, Factory brings it up #143) a sample service into a real dispatch namespace.
2. Run a load profile (e.g. 50 VUs, 30s) against the deployed endpoint → assert it completes and emits evidence JSON with real numbers.
3. Pass case: a lenient SLO (e.g. p95 < 2s, errors < 1%) against a healthy service → gate returns pass.
4. Fail case: a strict SLO (e.g. p95 < 1ms) OR load against a deliberately throttled/broken service → gate returns fail with the violated metric. (Proves the gate actually gates — not always-green.)
5. Tear down.

Deliverable

• test/e2e/load-harness.e2e.ts running a real load run (Container/CI k6) against a deployed Cloudflare environment with both a passing and a failing SLO assertion, one command in CI.

Anchor files

• new src/environments/load-harness.ts (+ k6 Container/CI runner)
• src/ports/environment.ts / src/environments/* (from EnvironmentProvider port + Cloudflare ephemeral-environment provider (Workers for Platforms + Containers); K8s/Daytona optional #142/Verification-stack descriptor contract + deployer: repos declare their full stack, Factory brings it up #143)
• src/observability/* (evidence shape precedent), src/index.ts

Fixes #144

Summary by cubic

Adds an in-cluster load harness that runs k6 as a Kubernetes Job, checks SLOs, and emits machine-readable evidence. Hardened to fail closed on zero-traffic and validate evidence integrity; this gates deployed stacks and fulfills #144 in the Factory Verification & Environments epic.

• New Features

  • runLoad(environment, profile) creates a k6 Job, parses results, applies SLOs, and returns pass/fail with measurements and evidence JSON (factory.load.evidence.v1).
  • Declarative profiles (JSON/YAML via loadLoadProfile(path)): targets (endpoint or URL), VU/RPS with optional ramp, thresholds, and latency histogram; arrival-rate and VU scenarios supported.
  • Evidence validation: strict schema checks (errorRate vs counts, latency ordering, cumulative histogram integrity); zero-traffic runs add a requestCount violation.
  • Kubernetes integration: KubectlLoadJobClient, hardened Job/ConfigMap with pinned DEFAULT_K6_IMAGE, and renderK6Script.
  • E2E + CI: kind-based test/e2e/load-harness.e2e.ts deploys a sample service, runs pass/fail SLOs, writes artifacts/load-e2e/*.json; GitHub Actions job creates a kind cluster, pins the kubectl context for the run, and uploads evidence artifacts bound to the head SHA.
  • Docs: docs/load-harness.md with usage and profile examples.

• Dependencies

  • Dev: add tsx and test:e2e:load script.

^{Written for commit 94d2454. Summary will update on new commits.}