PR benchmark guard (base vs head)

[AnshulPatil2005]

Add a lightweight PR benchmark workflow that runs base and head on the same runner and posts non-blocking regression warnings per case.
this idea is from Part 2 Task 1 of (opencax/GSoC#114 (comment))

Proposed approach

• integrate into existing .github/workflows/manifold.yml
• compare base vs head on the same runner using separate git worktrees
• use extras/perfTest as the initial benchmark target
• run 3 repeats for each side
• compare median of run means
• non-blocking warning threshold:
  • >= 20% and >= 10 ms
• upload raw benchmark artifacts and add a markdown summary in job output

Confirmation needed

1. Is perfTest the right initial target, or should I switch to another benchmark entry point?
2. Are the initial thresholds (20% + 10ms) acceptable for first rollout?
3. Keep this lane non-blocking for now, then tighten later after data collection?

does this look good ?

[elalish]

That all sounds like a reasonable place to start. Let's see how this goes and then decide how to adjust as necessary once it's running. I have the feeling we won't ever make this blocking, due to too much of it being out of our control, but it'll be nice to have the data.

[AnshulPatil2005]

Yes. On second thought, keeping it non-blocking makes more sense. I’ll proceed with it and then i can tweak it according to the data received