macOS arm64 SYN evidence for the adaptive-paper engine and mover trial mode

[seathatflowsinourveins]

Summary

First macOS arm64 run of the us-equities paper engine's synthetic suites, recorded with scripts/host_receipts.py record (evidence class local_integration, stage use, platform macos-arm64, second physical machine, result pass). No engine or blueprint source changed; nothing touches a broker, credentials, market data or study data.

• Host receipt: evidence/hosts/macos-m5pro-20260924/macos-m5pro-20260924--nautilus-trader--use--20260924.json
• Artifacts in evidence/artifacts/macos-syn-e2e-20260924/: the hash lock and its input, the invented-ticker mover scan, the mover SYN receipt, the capacity fixture output, and summary.json (host facts, per-module counts, the sandboxed first attempt, SYN detail)
• Registries: manifests/evidence.json registers the 7 new files and re-hashes .github/osv-scanner-lockfiles.json. That inventory now lists the new lock for OSV (its .in is covered_by_lockfile), as tests/test_osv_lockfile_coverage.py requires. The OSV API returned 0 advisories for the 20 pins on 2026-09-24.
• component_matrix.py --write and new_host_grand_list.py --write produced no change (nautilus-trader is not a landscape winner, so no platform status moves).

Host and runtime

Apple M5 Pro (Mac17,8), 18 cores (6 Super, 12 Performance), 24 GB, macOS 26.5.1 (25F80), arm64.
CPython 3.12.14 (uv 0.12.18, installed with --no-bin into the worktree's gitignored .runtime/). uv pip sync --require-hashes installed 20 packages, including the lane pins nautilus-trader 2.0.0rc5 (cp312-macosx_11_0_arm64 wheel) and alpaca-py 0.44.0, plus the engine recipe's numpy 2.5.3 and pandas 3.0.6. The installed set matches the lock exactly, and uv pip check is clean.

Recorded results

Command (repository root, pinned venv)	Exit	Result	Seconds
python -m unittest discover -s tests -p 'test_adaptive_paper_*.py' -v	0	732 run, OK (skipped=1)	112.3
... -p 'test_native_faults_min.py' (native-faults offline cases)	0	17 OK	0.15
... -p 'test_adaptive_market_research.py'	0	28 OK	0.30
mover_runner.py synthetic --scan <artifact scan> --allow-stale-scan	0	passed, SYN: 2 entries, 2 X2 exits, 4 native fills, 0 rejections, flat, cash_match, no recovery hand-off	4.6
benchmark.py	0	passed: 180 submissions, 180 native fills, 90 round trips, peak 179 per 60 s, 0 broker connections, 0 paper orders, flat, reconciled	60.6

The skip is by design: exchange_calendars is not in the lane runtime (test_adaptive_paper_sessions). The mover P&L comes from mover_runner.synthetic_path's scripted prices, so it is not an outcome.

Sandboxed first attempt. In Claude Code's Seatbelt sandbox, the same adaptive-paper command ran 732 tests with 3 errors, all PermissionError: [Errno 1] Operation not permitted from socket.bind in MetricsHttpServerTests. The sandbox denies local port binding. The rerun, and the recording, ran with the sandbox disabled. Every other command passed inside the sandbox, where ~/.config/native-agent-stack was unreadable. ~/.local/state/native-agent-stack (engine STOP, locks, ledgers) did not exist before the runs or after them.

Limits

• This is local_integration only: the real NautilusTrader LiveNode engine ran against the repository's synthetic ports. It is not paper or live acceptance for Alpaca or IBKR.
• Not run: the mover-early-entry study tests, the live native-faults harness, the full suite in the pinned runtime, and the 42 unchanged upstream adapter tests.
• The trading-nautilus profile lists bwrap, which macOS lacks, so the network-namespaced engine replay is not covered. No hardware_profile.py --record-host entry was made; this host has 24 GB, not the 64 GB of the projected macOS host.
• The receipt carries only the recorder's self review; an independent review is still open.

Checks

validate.py, host_receipts.py validate, component_matrix.py --check, new_host_grand_list.py --check, evidence_manifest.py --check, validate_catalogs.py, validate_foundation.py, landscape.py, validate_convergence.py --all-recorded, build_verdicts.py --check and build_ecosystem.py --check all passed. With the files staged, the OSV coverage, host receipt, platform status and staleness tests passed (166 OK). gitleaks dir on both new directories and on the two modified files found no leaks.

The full python3 -m unittest on this host (system Python 3.14.7, inside the sandbox) ran 3932 tests with 2 failures, 36 errors and 510 skips. I reran the 8 affected modules (350 tests) without the sandbox. That cleared every sandbox error (loopback bind, hard-coded /tmp, ps) and the OSV coverage failure, which was a race: I edited the inventory while the suite ran. Three failures remain, all unrelated to this change:

• test_adoption_bootstrap_macos.LlamaWrapperQuotingTests.test_install_path_with_shell_syntax_is_data_in_the_wrapper is macOS-specific. SIP strips DYLD_LIBRARY_PATH from the test's #!/bin/sh fake server, so the printed value is empty. It passes on Linux runners.
• The two test_catalog_freshness_propose setUpClass errors happen because this host's global gitleaks pre-commit hook (core.hooksPath) blocks the scratch-copy commit on findings in existing catalog files. CI has no such hook.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, add credits to your account and enable them for code reviews in your settings.