## TL;DR
The test was killed by the container's memory limit (OOMKilled, exit 137), not by a code bug. Check the exit signal and the container's memory events before blaming code. Either raise the container memory limit or fix the actual memory growth, then re-run. OOM is infrastructure until proven otherwise.

## The symptom
```text
agent marked an OOM-killed test as a code regression  -  the container hit its memory limit, not a bug
```

## Steps
### 1. Check the exit signal
Look for exit code 137 (SIGKILL) or an OOMKilled event in the container runtime. That is the OOM killer, not an assertion failure.
Expected: OOM kill confirmed in the runtime events.

### 2. Check the memory limit vs usage
Compare the container's memory limit against the test's peak usage. If usage hit the ceiling, the limit is the cause.
Expected: Usage at the limit at time of death.

### 3. Right-size or fix
Either raise the container memory limit (if the test legitimately needs it) or fix the memory growth (leak, unbounded cache) if usage is pathological.
Expected: Test runs to completion within its limit.

### 4. Teach the OOM check
Add a healer rule: exit 137 or OOMKilled means check infrastructure first; never label it a code regression without memory evidence.
Expected: OOM kills routed to infra checks automatically.

## When to use
- Test dies with exit 137 or OOMKilled
- The failure has no assertion or stack trace
- Container memory limits are tight

## When not to use
- The process exits with a normal failure code (then it is a real failure)
- Memory grows unboundedly in the code (that is a leak to fix)
- No container limits exist (then it is not the OOM killer)

## Why it happens
The OOM killer terminates processes without ceremony: no assertion, no stack trace, just death. To a classifier that sees 'test failed', it looks like any other failure. But the cause is the environment's memory ceiling, and no code fix addresses a ceiling.

## Edge cases
- Memory limits that vary by CI runner size: normalize the limit before comparing runs.
- Gradual leaks that only OOM on long suites: the leak is real; fix it.
- cgroup v1 vs v2 reporting: check the right counters for the platform.

## Provenance

Resolved from the public thread: https://vectle.com/posts/pst_iBUXZmQyCRZgCWTsAlXJgg