Epic: Async Scheduling Resource Metadata, Admission, and Observability

Priority Level

High

Epic Scope

Track implementation and quality gates for the async scheduling architecture. This issue is no longer the architecture specification; the plan files are the source of truth.

Source of Truth

Architecture, UML alignment, contracts, terminology, and cross-cutting decisions live in PR #658 under plans/645/, especially:

• plans/645/README.md
• plans/645/architecture.md
• plans/645/contracts.md
• plans/645/task-admission.md
• plans/645/request-admission.md
• plans/645/capacity-model.md
• plans/645/observability.md
• plans/645/benchmark-plan.md
• plans/645/migration-and-cleanup.md
• plans/645/issue-map.md

If this issue and plans/645 disagree, update the plan first and keep this issue as implementation tracking only.

Implementation Lanes

• Metadata lane: #641 -> #646 -> #653 -> #652.
• Scheduler admission lane: #646 -> #644 -> #647.
• Request admission lane: #644 -> #654 -> #657 -> #635.
• Correlation lane: #657 + #635 + #647 -> #648.
• Benchmark harness lane: #644 + #654 -> #649.
• Documentation lane: #652 + #654 + #657 + #648 + #649 -> #660.
• Policy lane: #649 -> #650 -> #651.

Epic PRs should target epic/645-async-scheduling while the epic branch is active unless the PR explicitly documents a different target.

Epic Quality Gates

• Generator/plugin metadata lands behind the SchedulingMetadata contract and does not expose scheduler internals.
• Scheduler grouping consumes resolved metadata through the plan-defined bridge, and the legacy hint resolver path is removed or collapsed with no independent introspection fallback.
• Task admission uses an explicit lease boundary before worker spawn, with root/from-scratch and downstream work visible to the same admission path.
• Request admission is a distinct model-call-time layer with request leases, exact-once release, adaptive provider/model/domain state, and no durable Throttle* production names.
• Capacity vocabulary and snapshots distinguish engine, row-group, task-admission, request-admission, static provider caps, adaptive request caps, and transport/session pools.
• Scheduler/request telemetry remains generic and independently useful, with correlation added through the plan-defined read-only bridge.
• #649 produces reusable benchmark artifacts before follow-on policy work relies on measured fairness/throughput claims.
• Final docs explain the two-stage model and capacity/observability surfaces without reintroducing removed names as current APIs.
• Repo search gates pass for stale production/current-doc terminology called out in plans/645/migration-and-cleanup.md.
• Existing async scheduler, fair queue, request-admission, end-to-end async engine, docs, and recipe validations continue to pass as applicable to each PR.