feat(batch): real ECS-backed SubmitJob execution (batch 2)

[vieiralucas]

Summary

The headline differentiator for the cycle-5 winner: AWS Batch SubmitJob runs a REAL container, not a fake. Every other free emulator fakes Batch compute (MiniStack jumps straight to SUCCEEDED with no container; LocalStack paywalls it).

• Wire ECS state + runtime into BatchService (with_ecs). SubmitJob resolves the job definition's containerProperties (+ this submit's containerOverrides), ensures a fakecloud-batch ECS cluster, RegisterTaskDefinition (mapping vcpus/memory — legacy fields or resourceRequirements — to ECS cpu/memory), and RunTask — reusing all of ECS's real container / portability / k8s handling through its public handle() API (mirrors how autoscaling drives EC2).
• A background poller maps the ECS task's real lifecycle + container exit code onto the Batch job: SUBMITTED → STARTING → RUNNING → SUCCEEDED (exit 0) or FAILED (carrying the real container.exitCode). Launch is non-blocking (ECS RunTask backgrounds the container pull/start → no client timeout).
• NO auto-success: with no container runtime wired the job stays SUBMITTED honestly — exactly the rival anti-pattern Batch beats.

Test plan

• Unit (10): resource mapping from both shapes, job-def latest-revision + override resolution, no-runtime-stays-SUBMITTED, control-plane lifecycle.
• Docker-gated e2e batch_real_execution.rs: real alpine container — exit 0 → SUCCEEDED/container.exitCode 0, exit 7 → FAILED/exitCode 7. Both PASS locally with Docker.
• Control-plane e2e batch.rs still passes; cargo clippy -p fakecloud-batch --bin fakecloud --all-targets -- -D warnings clean; doc-counts passes.

Surface

Docs page updated to reflect real execution. The batch_real_execution e2e mirrors the EC2/ECS Docker-gated pattern (panics in CI without Docker, skips locally).

Next batches

Array jobs + dependsOn + retry/timeout, then a CloudFormation AWS::Batch::* provisioner + tfacc.

---

Summary by cubic

SubmitJob now runs real containers on ECS via fakecloud-ecs, advancing jobs through STARTING/RUNNING and finishing as SUCCEEDED or FAILED from the container’s exit code; without a container runtime they remain SUBMITTED.

• New Features

  • Resolve container properties + overrides, ensure a fakecloud-batch ECS cluster, register a task definition (map legacy vcpus/memory or resourceRequirements → ECS cpu/memory), then run a task.
  • Background poller syncs ECS task state to job status and records container.exitCode; launch is non-blocking.
  • No auto-success if ECS isn’t available.

• Refactors

  • Docker-gated e2e now uses resourceRequirements instead of deprecated vcpus/memory; docs updated.

^{Written for commit 97ca700. Summary will update on new commits.}