Build async scheduling benchmark harness for admission and fairness work

[eric-tramel]

Priority Level

High

Epic: #645
Depends on: #644, #654
Uses: #647 and #648 when available
Feeds: #650, #651
Related: #657, #635
Target branch: epic/645-async-scheduling while the epic is active.

Source of Truth

Specification details live in PR #658 under plans/645/, especially benchmark-plan.md, capacity-model.md, observability.md, task-admission.md, request-admission.md, and issue-map.md. This issue tracks the benchmark harness slice and gates only.

Implementation Scope

Build the reusable async scheduling benchmark harness and artifact schema.

This issue owns:

• a repo-local benchmark entrypoint with baseline/candidate comparison support;
• machine-readable JSON/CSV artifacts plus a concise markdown summary;
• capacity-plan capture, observed maxima, run metadata, command lines, commit SHAs, deterministic hashes where applicable, and runtime/machine info;
• CI-smoke mode using deterministic mock endpoints;
• required scenario families from plans/645/benchmark-plan.md: queue/admission microbench, heavy-root downstream, hidden-waiter proof, idle/utilization proxy, end-to-end A/B timing, dynamic request-call workloads, and correlated trace output when available.

This harness does not replace correctness tests and does not make the full benchmark matrix a mandatory CI perf gate.

Quality Gates

• Harness runs against two refs/checkouts without manual editing.
• Artifacts are stable enough for Implement TaskAdmissionController lease boundary for async scheduler #644, Add bounded-borrow admission policy for heavy-root async workloads #650, Design resource-vector and provider-aware admission policies #651, observability, and docs evidence.
• Smoke command writes machine-readable artifacts without live providers.
• Neutral scenarios report no worse than the plan-defined threshold unless the PR explicitly documents an accepted fairness/throughput tradeoff.
• Heavy-root and hidden-waiter scenarios report the required lag/debt/waiter metrics even for PRs not expected to improve them.
• Output equality is checked for deterministic scenarios.

Validation

Run the CI-smoke benchmark, at least one local two-ref comparison, artifact schema validation, deterministic hash validation, and make check for the benchmark code.