[Accepted risk] Enforce bounded-worker memory containment on macOS #510

Risk

Issue #301 adds pre-import Unix resource caps, but Darwin rejects every tested finite RLIMIT_AS value with ValueError: current limit exceeds maximum limit. The harness therefore records address_space_bytes as unsupported on macOS while still enforcing file-size, open-file, real-user process, timeout, output-retention, and artifact-budget bounds.

This is an explicit accepted risk, not evidence that memory is bounded on macOS.

Owner and expiry

• Owner: maintainers
• Review by: 2026-08-07
• Priority: P2

Exit criteria

• Select a primary-source-backed macOS containment mechanism that works before heavyweight worker imports.
• Add a deterministic regression proving a worker cannot exceed the selected memory ceiling.
• Record requested, effective, and unsupported memory containment in run and per-case evidence.
• Keep unsupported/failed setup non-rankable under the lifecycle contract.
• Run exact-SHA CI and resolve every review thread before closeout.

Evidence

• Python documents resource as Unix-only and resource availability as system-dependent: https://docs.python.org/3.12/library/resource.html
• Local aarch64-darwin probes rejected finite RLIMIT_AS values from 8 GiB through 256 GiB; RLIMIT_FSIZE, RLIMIT_NOFILE, and RLIMIT_NPROC were accepted.