fix(backend): enable faulthandler so the SIGABRT crash-loop is diagnosable (#12777)

[mrveiss]

Thinking Path

autobot-backend aborts with SIGABRT every ~25–55 minutes and produces zero diagnostics. The issue already ruled out every ordinary cause from cgroup counters — memory.events all-zero with a 3.4 GiB peak against a 12 GiB cap, pids.peak 65 against a 512 limit — so this is a native abort. And a native abort is precisely what the current setup cannot record.

Two things independently destroy the forensic trail:

1. faulthandler was never enabled — no PYTHONFAULTHANDLER, no -X faulthandler. A fatal signal inside a C extension unwinds without printing a Python stack, so backend-error.log stops mid-stream and resumes at the next boot.
2. Core capture is broken on the platform — core_pattern pipes to /wsl-capture-crash, which does not exist. systemd reports code=dumped while the core is discarded.

So the first fix is not a fix for the crash — it is making the crash observable. Without a stack, step 3 of the issue ("pin the aborting extension") is unactionable.

What Changed

One line on the canonical unit template (authoritative per #11670):

Environment="PYTHONFAULTHANDLER=1"

The next abort now writes a native + Python traceback to StandardError, which the same unit already appends to backend-error.log.

Deliberately not included

• core_pattern is host/WSL configuration, not repo state. The issue itself asks for it to be either fixed or explicitly opted out of with a recorded reason — that's an owner call, not a drive-by.
• Pinning or rebuilding the suspect extension (uvloop / chromadb / torch-onnx against a very new Python 3.14) stays a hypothesis until a stack exists. Worth being precise: the single uvloop OSError: [Errno 88] Socket operation on non-socket appears once in the whole log and preceded only one of the two observed aborts — it cannot explain both, and acting on it now would be guessing.

Verification

$ python3 -m pytest tests/test_backend_service_faulthandler_12777.py -q
3 passed

$ python3 -m pytest tests/ -q
976 passed, 1 skipped

Three render tests, following the test_pg_hba_multiuser.py precedent (#10636): faulthandler stays enabled; stderr still has somewhere to land (the two only work as a pair — enabling faulthandler while stderr goes nowhere leaves the crash just as invisible); and Restart=always survives, since the loop is only survivable because systemd restarts the unit.

Mutation-verified — deleting the env line fails the first test:

env line removed  → test_faulthandler_is_enabled FAILED
env line restored → 3 passed

The test registers Ansible's dirname/basename filters so the real template renders outside Ansible, rather than testing a copy.

Deployment note

This is a systemd unit change, so it only takes effect after the backend role is redeployed — it will not reach the box via a code-only sync. The crash loop stays undiagnosable until that happens.

Model Used

Claude Opus 5 (1M context)

Closes #12777

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!