eval: service capacity scoring and sealed agent slice manifest

[ShivanshuPurohit]

Scoring path for the agent-goodput arena: service capacity at the contract boundary, the sealed slice manifest, and the retained per-turn row shape. Increment into eval-revamp-agentic-coding; Codex's execution path (bridge, worker loop, runner, commission, dashboard) is the consumer.

Consumer. qualification_runner calls capacity() per arm per window and service_verdict() once GP1's replay lands (Codex). Until then both modules are listed in scripts/island_baseline.txt; those two lines come out in the PR that wires the runner. The merge of the feature branch to main is gated on that consumer.

Casualty. resident_measurement, resident_schedule and the version 8–15 ladder in speed_verdict for the agent profile, plus the batch-cell surface they serve (screen family on the eval and chain sides, batch core, cell authority). Not deleted here: they still carry the Qwen arena, which runs from main until it retires (owner decision 2026-09-27: the GLM arena moves to this feature branch; Qwen deletion can go). Retirement trigger: the cutover tranche on this branch after GP1/GP2 land, sized in the private census at roughly −10,000 production and −7,000 test lines. No deletion was available in this change.

What it does.

• cacheon/eval/service_capacity.py: TurnRecord rows, fixed-work rate that refuses any work other than the sealed slice, attainment under a ServiceContract (decode floor, queue-inclusive TTFT bound, attainment fraction), capacity by log-linear interpolation of the attainment crossing between the two bracket loads, and the paired verdict through the existing invariant_decision. load_reads_jsonl reads the retained row shape.
• cacheon/eval/agent_slice.py: reopens a slice manifest over its directory of one-trace files, proving names, order, bytes and per-file session identity against the sealed digest; expected_work(load) is the per-root (main, inner) accounting the scorer checks.
• Tests replay the retained 2026-09-26 final budget-4 pairs (AIPerf profile exports converted to rows, 3,150 turns): the step rule scores the winner at +2.8%, interpolation at the locked contract (60 tok/s, 5 s, 0.8) gives a 1.116 capacity ratio, and a tighter contract makes the incumbent infeasible rather than a zero denominator.

Ratchets. cacheon ceiling +465, tests +323, scripts +2 (island entries), all in this diff. surface-change: modules — two new modules with their __all__.

Docs. No page changes here: the evidence shape and manifest are settled in the private GP0 contract and land in docs/ with the feature PR that adds the arena; the module docstrings carry the invariants until then.

Net lines. Production +465 / −0. Tests +323 / −0 plus a 142 KB gzipped fixture. Baselines +8 / −6.

🤖 Generated with Claude Code