LPX-649: extract queue & scheduler into their own ECS services

[stevethomas]

Hey, I made a thing! 🥳

LPX-649 — extract queue & scheduler into their own ECS services.

What problems are you solving?

YOLO bundles web + queue + scheduler into one Fargate task, coupling three workloads with different scaling shapes onto a single desiredCount. This makes each a service it can scale independently:

• web — target tracking (unchanged).
• queue — its own service, scale-to-zero by default. Backlog-per-task target tracking (ApproximateNumberOfMessagesVisible / RunningTaskCount via CloudWatch metric math, no Lambda), plus a step-scaling alarm that lifts it 0→1 the instant a message lands (target tracking can't divide by zero running tasks). Opt-in Fargate Spot (~70% cheaper). Costs ~$0 idle.
• scheduler — its own pinned-singleton service (min=max=1, never a scalable target), deployed stop-then-start so a rollout never briefly runs two crons. Drops the ->onOneServer() requirement.

Topology is encoded by location, not a flag (your call on the design):

Manifest	Means
tasks.web.queue / tasks.web.scheduler	bundled in the web container — warm, instant pickup. Unchanged.
top-level tasks.queue / tasks.scheduler	extracted into its own service with the grown-up config.
both, for one workload	hard error — pick one.

So tasks.web.queue = "a chore the web box also does"; tasks.queue = "a workload that stands on its own." Nothing breaks — existing manifests are untouched, extraction is additive opt-in.

Also in this PR:

• --group on deploy / run; scale --queue (min 0 = scale to zero); group-aware DeployerPolicy.
• One image serves every role — the task-def passes the role as the container command and the entrypoint dispatches with a per-role graceful drain.
• Retires the dead EC2-era RunsOnAws*Environment detectors + the unused ParsesOnlyOption concern (no Fargate implementors).
• Hardened Manifest::put's surgical YAML writer to fall back to a full re-dump for an inline-empty-map parent (queue: {}) rather than corrupting it.
• Docs: scaling guide, manifest + commands reference, yolo.yml / Dockerfile stubs.

Is there anything the reviewer needs to know to deploy this?

• No infra was touched and nothing was merged — this is code + docs only.
• No breaking manifest change. Bundled tasks.web.queue/scheduler keep working exactly as before; CL's yolo.yml needs no changes. Extraction is opt-in.
• Multi-tenancy: a standalone queue is one service per app on the default/landlord queue; per-tenant queue fan-out is out of scope and composes with LPX-601.
• The web ALB health-gate is unchanged; a --group-scoped deploy that omits web skips that wait and relies on the ECS circuit breaker for the headless services.
• 553 Pest pass · PHPStan clean · Pint clean · VitePress docs build clean. Rebased onto latest main (incl. feat(sync): heartbeat + realistic timeout for slow AWS waiters #72 and the IAM-policy-drift / elasticache deployer changes).

🤖 Generated with Claude Code