Add kernel warning system for overflow detection

[adenzler-nvidia]

Summary

Adds a structured warning system for detecting and reporting overflow conditions in kernels. Previously, wp.printf calls went to stdout and cluttered output. Now warnings are captured via atomic flags and emitted to stderr with actionable messages.

Changes

New Warning System

• WarningType enum for categorizing warnings: NEFC_OVERFLOW, BROADPHASE_OVERFLOW, NARROWPHASE_OVERFLOW, CONTACT_MATCH_OVERFLOW, GJK_ITERATIONS, EPA_HORIZON, HFIELD_OVERFLOW
• check_warnings(d) - reads flags from GPU and emits to stderr with suggested values
• get_warnings(d) / clear_warnings(d) - utility functions for programmatic access
• Warning arrays added to Data: warning (flags) and warning_info (suggested values)

Kernel Changes

• All overflow wp.printf calls are now guarded by wp.static(m.opt.warning_printf) for compile-time removal
• Kernels set warning flags via wp.atomic_max with suggested parameter values
• Affected: forward.py, sensor.py, collision_convex.py, collision_gjk.py

New Option

• opt.warning_printf (default: True) - controls whether printf warnings are compiled into kernels
• testspeed and viewer disable this and use check_warnings() instead

Cleanup

• Removed dead wp.printf calls in smooth.py (unreachable due to io.py validation)

API

Warnings are emitted per-step to stderr:

Warning: nefc overflow - increase njmax to 32
Warning: nefc overflow - increase njmax to 36

Programmatic access:

m.opt.warning_printf = False  # disable printf, use check_warnings instead
mjw.step(m, d)
mjw.check_warnings(d)

Testing

New warning_test.py with 8 tests covering initialization, detection, clearing, graph capture scenarios, and the printf option.

[erikfrey]

Nice!

I wouldn't be against defaulting warning_printf to false if we want to discourage this for general use. WDYT @adenzler-nvidia ?

[adenzler-nvidia]

I'm torn on that one - I think before we had the printfs we had a lot of silent overflows. Any user not opting in to this warning system would be in that situation again.. So I think it's safer to default to True and give everyone using this warning handler or implementing it themselves the option to turn off the printfs.

[thowell]

the overflow warnings are very helpful for debugging and my vote would be to keep them on by default. we could note in the the performance section of the documentation to set this option field to False for best performance?

[thowell]

@adenzler-nvidia is it possible for us to utilize this (in combination with some sort of option) to error / stop simulation when a warning is detected?

[adenzler-nvidia]

The error handler can definitely be expended to do something like this. I would say it's up to the application integrating MjWarp, I don't think this should be hardcoded into MjWarp.

I can see an app like MjLab implementing their own error handler to do this. We can do it for the renderer/testspeed as well.

This is also the reason I vote for keeping the printfs in by default - it's a good thing for anyone integrating MjWarp. Once you need more than that, you can turn off the prints and implement an error handler.