Make the macOS bootstrap tests pass on a SIP-enabled Mac without npm (tests only; supersedes #178)

[seathatflowsinourveins]

Summary

On a developer Mac with System Integrity Protection (SIP) enabled and no npm on PATH, tests/test_adoption_bootstrap_macos.py has one failure and one error on unmodified main. Both are test-harness problems, so this PR changes tests only. adoption/bootstrap-macos.sh is unchanged.

Reproduced at d478fd5 on macOS 26.5.1 (25F80) arm64. csrutil status reports enabled, and ls -lO /bin/sh shows restricted. The interpreter is uv-managed CPython 3.13.15. Temp paths are shortened to <tmp>:

FAIL:  LlamaWrapperQuotingTests.test_install_path_with_shell_syntax_is_data_in_the_wrapper
       AssertionError: Lists differ: ['<tmp>/…/llama-server', ''] != ['<tmp>/…/llama-server', '<tmp>/…/llama-cpp-b1']
ERROR: PlatformDependencyInstallTests.test_places_the_verified_tarball_correctly_when_the_prefix_is_reached_through_a_symlink
       AttributeError: 'PlatformDependencyInstallTests' object has no attribute 'verified_sha256'
Ran 116 tests: FAILED (failures=1, errors=1, skipped=37)

1. Llama wrapper quoting test under SIP (commit carried over from #178)

• Cause: the stand-in llama-server was a #!/bin/sh script. With SIP enabled, dyld purges DYLD_* variables when it launches the protected /bin/sh, so the stand-in could not see the wrapper's DYLD_LIBRARY_PATH export. The real llama-server is an unprotected Mach-O, so the generated wrapper is correct.
• Change: this is Make the llama wrapper quoting test pass on SIP-enabled Macs (wrapper unaffected) #178's commit 3fd39cd, cherry-picked unchanged with -x. Neither the test file nor bootstrap-macos.sh has changed on main since Make the llama wrapper quoting test pass on SIP-enabled Macs (wrapper unaffected) #178's base (4911baf), and the commit applied cleanly.
  • The stand-in is now a script whose shebang is a symlink to sys.executable in the test's temp dir. The symlink keeps spaces out of the shebang. The script prints sys.argv[0] and os.environ.get('DYLD_LIBRARY_PATH', '').
  • The PWNED/PWNED2 assertions are unchanged. So are the exec-target ($0) check and the exact [target, prefix] assertion.
  • Make the llama wrapper quoting test pass on SIP-enabled Macs (wrapper unaffected) #178 also adds two things:
    • A control run: if the test's own interpreter drops a DYLD_* value handed to it directly (for example the SIP-protected /usr/bin/python3), only the DYLD_LIBRARY_PATH assertion is skipped, and only after the other assertions have run.
    • The wrapper now runs with DYLD_LIBRARY_PATH unset instead of empty, so a missing export line fails on every host.
• Make the llama wrapper quoting test pass on SIP-enabled Macs (wrapper unaffected) #178's evidence still applies because the inputs are identical. That evidence covers the wrapper against the pinned b11057 asset, the mutation checks, and its CI runs.

2. Missing npm skip

• test_places_the_verified_tarball_correctly_when_the_prefix_is_reached_through_a_symlink uses the npm-packed fixtures. It reads self.verified_tarball directly and self.verified_sha256 through _platform_dependency_pin. setUpClass builds these only when npm is on PATH.
• Its npm-dependent siblings carry @unittest.skipUnless(NPM, "native npm unavailable"), but this test was added in macOS arm64 clean install: Homebrew prerequisites, darwin pins, launchd agents (brew-services semantics), embedding acceptance, recording-tooling CI #94 without it. The test passed on green main CI, so hosted runners have npm and it will keep running there.
• Change: add the decorator. The class's other three undecorated tests use no npm fixtures and still run without npm.

Checks on this Mac

Run of tests.test_adoption_bootstrap_macos	Result
Unmodified main, npm absent	116 tests: failures=1, errors=1, skipped=37
This branch, npm absent	116 tests OK, skipped=38. The llama test ran and passed; the symlink test skipped with native npm unavailable
This branch, npm present (mise-managed node 24.21.0 / npm 11.19.0 prepended to PATH)	116 tests OK, skipped=16. Both target tests ran and passed

The remaining skips are environment gaps: PyYAML not installed (14), shellcheck not installed (1), the non-macOS guard test (1), and the npm tests (22, only when npm is absent).

The branch was rebased onto 636751a (#224), which touches none of these files. After the rebase:

• Validation: python3 scripts/validate.py (system Python 3.9.6) passed with 69 components, 5338 hashed files, 4 profiles and 145 receipts.
• Module: tests.test_adoption_bootstrap_macos was re-run: OK, skipped=38.
• Full suite (python3.13 -m unittest, npm absent, at this PR's head): ran 4714 tests with failures=1 and skipped=647.
  • The one failure is tests.test_new_host_grand_list.RecordHostToGrandListTests.test_recorded_host_gets_non_null_measured_in_the_grand_list. It fails the same way on unmodified main (d478fd5, re-run in a clean checkout) on this host, so it is unrelated to this PR.
  • Cause: a host-specific false positive in the local-user-name guard makes hardware_profile.py --record-host refuse the test's fixture id (exit 2).
  • A run before the rebase, on d478fd5 plus these two commits, gave the same single failure (4713 tests).
• Evidence: manifests/evidence.json re-hashes only tests/test_adoption_bootstrap_macos.py, through host_receipts.register_file.

Relation to #178

This PR contains #178's commit plus the npm fix, re-hashed once. If this merges, #178 is redundant and can be closed. It has been left open for you to decide.

Evidence class: local integration on one SIP-enabled Mac. No product change, no receipt added, and no launchd, embedding or model run.

🤖 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.