Implement multi-platform CI sync testing

[bar-discord-issue-bot]

Task

Create a GitHub Actions CI workflow that runs engine sync tests on both arm64 and amd64 architectures, and compares the resulting checksums against each other and against historical "golden" runs to automatically detect sync regressions.

Context

Currently, verifying that the engine remains deterministic across different platforms requires manual testing. As we look to introduce new compiler optimizations (like enabling SSE4) or refactor core simulation code, we need automated safeguards to ensure we don't introduce desyncs between players on different hardware.

The proposed solution is to run sync tests as part of the CI pipeline on both amd64 and arm64 runners. A dependent job will then compare the outputs to ensure cross-platform determinism. However, comparing architectures against each other isn't enough, as a bug could theoretically alter the simulation deterministically across all platforms at once. To catch this, the CI must also compare the results against historical or "golden" checksums from previous revisions. We can draw inspiration from the BAR game repository's integration tests, which write results to disk in one job and evaluate them in another, potentially storing a rolling history of recent runs in a structured format.

Acceptance Criteria

• GitHub Actions workflow is updated to run sync tests in separate jobs for both amd64 and arm64 runners.
• A dependent CI job is created to collect and compare the checksum outputs from both architectures, failing the build if they diverge.
• Test results are saved in a structured format (e.g., JSON or CSV containing PR/commit hash, contributor, test results, and timestamp) to maintain a history of recent runs.
• The comparison job evaluates current checksums against historical/golden checksums to detect cross-revision sync breakages.
• The system includes a documented mechanism to update the golden checksums when intentional changes to the simulation or sync tests are made.

---

Author: Bruno D
Discord: https://discord.com/channels/549281623154229250/724924957074915358/1484690181482221638
Model: gemini-3-pro-preview

[Bruno-DaSilva]

@p2004a suppose we wanted to store the checksums (perhaps just a final hash of all the checksums) on every mainline commit, so we can compare a new build against the history of checksums. In other words, so main branch gets a "golden checksum" and if we do a sync test on a PR, we can flag if the sync testing diverges.

One idea was to store the "golden checksums" in a bucket somewhere, but that involves us setting up github <-> GCS. We can do that with workload identity federation (or storing a SA json but that's probably worse) but I'd rather not add that complexity if there's a better way.

Do you have any ideas that would be simpler? Really all we need is a simple key/value store. Or if we could somehow use github as the store on each commit...

[Bruno-DaSilva]

D'oh, what am I thinking. Action runs have artifacts, so we can just store the hash in artifacts and read from it. Yes, retention isn't infinite, but we only need the recent engine hash outputs so that's fine.