fix: diagnose TruthBench generated-code sandbox crashes

[kevincostner17]

Summary

The scheduled CI truthbench job fails intermittently (09-01, 09-04, 09-06) with:

REGRESSION: regression: gate generated_code_sandbox passed in baseline, fails now

The downloaded truthbench-results artifact gives the only evidence:

{"name": "generated_code_sandbox", "failures": ["generated code exited -11: "]}

That is a segfault (SIGSEGV) in the sandbox subprocess with empty stderr. It is not a timeout or a canary leak, and neither the run nor the report shows which generated script crashed.

This PR is the diagnose-first step: it doesn't change what passes or fails, but the next crash will be explained.

Changes

• benchmarks/truthbench/generated_code.py:
  • Run the sandbox with -X faulthandler, so a native crash writes its Python traceback to stderr. The dump holds only file, line and function names, and stderr is already scanned for canaries and redacted before it is reported.
  • Pin native thread pools and the locale in the otherwise empty sandbox environment (OMP_NUM_THREADS, OPENBLAS_NUM_THREADS, MKL_NUM_THREADS, POLARS_MAX_THREADS, RAYON_NUM_THREADS = 1; LANG=C.UTF-8). Without them, BLAS/OpenMP/Rayon size their pools from the host core count, which is the leading suspect for a crash that happens only on some runners.
• benchmarks/truthbench/runner.py: sandbox gate failures are prefixed with domain/surface, so the report names the crashing script.

If crashes continue after this, the faulthandler output should point at the module to fix. A narrowly scoped retry for signal exits (PR/scheduled ratchet only, never the release gate) stays a fallback, not part of this change.

Tests

tests/truthbench/test_generated_code.py uses the existing python= hook with a fake interpreter to test two things:

• The sandbox passes -I -X faulthandler and the pinned thread variables.
• A SIGSEGV crash yields a failure string containing exited -11 and the Segmentation fault faulthandler dump.

Results:

• pytest tests/truthbench: all pass
• ruff check benchmarks/truthbench tests/truthbench: clean
• make truthbench-pr: generated_code_sandbox: pass, overall: PASS

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: a6ca6935-716b-4ddf-ba31-bf6396c8fd62

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

FreshData benchmark report — performance

• freshdata: ?
• python: ?
• platform: ?

fixture	n_rows	n_cols	p50 s	p95 s	peak MB	repair %	false-repair %	preserve %	trust	monotonic	export %

Authored-code reduction (Metric 6)