Implement async capacity model and runtime snapshots

[eric-tramel]

Priority Level

High

Epic: #645
Depends on: #641, #646, #644
Related: #657, #635, #647, #648, #649, #660
Target branch: epic/645-async-scheduling while the epic is active.

Source of Truth

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

Implementation Scope

Implement and document the async capacity vocabulary and runtime capacity snapshots.

This issue owns:

• AsyncCapacityPlan or final plan-equivalent DTOs.
• Consistent names for engine selection, record windows, row-group execution, task admission, request admission, static provider/model caps, adaptive request caps, and transport/session pools.
• Runtime snapshots of configured capacity and observed maxima needed by telemetry and benchmarks.
• Alignment of scheduler/request-admission config language with the plan.
• Provisional evidence artifacts needed before Build async scheduling benchmark harness for admission and fairness work #649 is reusable.

This issue should not add public capacity knobs unless the plan and benchmark evidence justify them.

Quality Gates

• Capacity terminology matches plans/645/capacity-model.md and is used consistently by task admission, request admission, observability, benchmarks, and docs.
• Static provider/model caps are distinguished from adaptive request-domain limits and transport pools.
• Runtime snapshots can explain which layer is limiting progress without conflating task and request admission.
• No durable current API/docs terminology remains for removed throttle or LLM-wait names except intentional migration/history text.
• Tests cover capacity-plan construction, alias/static cap handling, observed maxima, and trace/benchmark artifact fields.

Validation

Run focused capacity-plan tests, affected scheduler/request-admission tests, docs checks if docs are touched, stale-term searches, and provisional benchmark artifact generation per plans/645/benchmark-plan.md.