Grade with the producing run's network and resource posture

[ppXD]

Summary

• Before this change, an acceptance setup step ran with AllowNetwork=true hard-coded (host network under bubblewrap), and neither setup nor check had cgroup ceilings. Benchmark and qualification grading ran the fixture's tests over the agent's workspace on the raw local runner.
  • Every lane now grades under an AcceptanceGradingPosture taken from the run whose bytes it executes: its network grant and egress allowlist, plus its tier's memory and CPU row, clamped by Sandbox:MaxAutonomy and narrowed by Sandbox:AgentMemoryCeilingMb.
  • AcceptanceGradingPosturePolicy.Bind wraps the grading runner once, narrow-only, for the setup and every oracle.
  • A missing posture fails closed to network off and the Confined ceilings.
• What a grade reports follows what the sandbox did, not what the spec asked.
  • Runners say which egress they enforce (ISandboxEgressEnforcement; LocalProcessRunner.EnforcedEgress).
  • A notice (setup network: …) heads the evidence only when the setup was actually severed or filtered.
  • A setup the sandbox severed fails as setup-failed-network-severed: …. That is infra (no revise round), and agent.code does not respawn it, because the posture comes from the same stored task every time.
• A check killed at the grade's memory ceiling (ResourceExhausted) is tests-resource-exhausted, class Environment, like tests-timed-out. Before, it was a Genuine tests-failed-exit-137.
• Lanes:
  • Executor branch, patch, multi-repo and local lanes: the run's own task.
  • Supervisor per-unit, baseline, captured, resolve (branch and patch) and branchless-stop lanes: the unit's stored task.
  • Stop over the integrated head: the run profile's tier.
  • BenchmarkRunner: its agent's task.
  • TaskLaunch benchmark cells: the posture all attempts agree on, otherwise fail-closed.

Owner-visible behaviour (setup network applies where bubblewrap confines; ceilings apply where Sandbox:CgroupRoot is delegated):

Producing run	Setup network	Setup + check ceilings
Confined	severed (was host network)	1024 MiB, 1 core (was uncapped)
Standard (default for AgentTask, agent.code, supervisor units)	severed (was host network)	4096 MiB, 4 cores (was uncapped)
Trusted / Unleashed	host network (unchanged)	6144 MiB, 4 cores (was uncapped)
Egress allowlist	only the operator's EgressAllowHosts, filtered where the host filters, severed where it cannot; no hosts → severed	the tier's row
Unknown producer / positional overload	severed	Confined row

• An operator setup that downloads on a network-off run no longer does where the sandbox confines. The fix is an egress allowlist or a higher tier.
• The operator stop floor over the integrated head now grades under the run profile's tier (default Standard) inside the 300 s default window. A CPU-heavy floor is likelier to hit tests-timed-out.
• Benchmark and qualification checks now run under the agent's own ceilings, and the TaskLaunch fallback is Confined.
• Baselines are memoized per posture.
• EvaluatorVersion is supervisor-acceptance/v9.

Test plan

• Unit:
  • derivation table
  • For() read through real configuration (Sandbox:MaxAutonomy, AgentMemoryCeilingMb)
  • EnforcedEgress table and notice table
  • severed / unconfined / filtered setup failures through the real grader + TestsPassGrader
  • a check OOM under Standard and Trusted postures
  • pinned literals (tests-resource-exhausted, setup-failed-network-severed:)
  • agent.code retry rows
  • BenchmarkTaskGrading posture and fail-closed
  • TaskLaunch GradedPosture
  • inventory of grade lanes and oracle-runner bindings
• Integration:
  • resolve posture Theory over branch and patch arms
  • executor flow under the real LocalProcessRunner
  • real BenchmarkRunner hands the oracle a runner bound to its agent's posture
  • supervisor per-unit fold flows
• Regression: full unit suite (11950 passed, 1 skipped); integration Acceptance|OracleRestore|NonCodingOracle|LocalAcceptance|Benchmark|Qualification|AgentNodeFlow|ReviseLoop (338 passed); .Supervisor|Oracle|PlanMap|UnderClaim|Resolve (679 passed)
• Mutation: Posture = null on the resolve patch arm fails 2 rows; For() without the host budget fails the configuration test
• Sandbox lane (Linux bwrap): AcceptanceGradingPostureE2ETests, covering a real severed setup that reports setup-failed-network-severed: and a host that reports the severance it enforces