[jallegri]

Summary

• Add a focused CI job that builds and installs the candidate wheel in an isolated environment.
• Reuse existing JSON, MessagePack, and Arrow metadata contracts, with guards against importing checkout source.
• Keep SDK runtime behavior unchanged. This checks the candidate wheel, not overall release readiness or published artifacts.

Related: Basekick-Labs/arc#726.

Test plan

• All 36 selected contract cases passed against the installed wheel.
• Independent review repeated all 36 cases from outside the checkout.
• Source-import guard rejected a deliberately injected checkout import.
• Helper Ruff lint/format, workflow YAML parsing, and diff checks passed.
• Hosted CI checks.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-04T02:58:41.320339Z	b23c388	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[xe-nvdk]

Thanks @jallegri. We broadened this in 0679577: the job now runs all of tests/unit against the installed wheel instead of 7 named query tests (out of 22 in those two files), and installs [dev,all] so the pandas and polars tests run instead of skipping. Your source-import guards are unchanged.

• Locally: 1121 passed against the wheel, with no arc_client module loaded from the checkout.
• Negative check: deleting arc_client/management/diagnostics.py from the installed package fails the job at collection (exit 2), so any module the wheel leaves out is caught, not only on the query path.
• Cost: about 45 seconds.

Renamed the script to scripts/check_wheel.py and the job to "Unit tests (installed wheel)".