Summary

test/e2e/test-runtime-overrides.sh test 14 ("Config unchanged after rejected override") crashes the container and aborts the test script. Introduced by #2659.

Root Cause

PR #2659 tightened apply_model_override() to require NEMOCLAW_MODEL_OVERRIDE or NEMOCLAW_INFERENCE_API_OVERRIDE to be set before triggering. Test 14 was updated to pass NEMOCLAW_MODEL_OVERRIDE=test alongside NEMOCLAW_CONTEXT_WINDOW=notanumber.

The validation logic in scripts/nemoclaw-start.sh does return 1 when NEMOCLAW_CONTEXT_WINDOW is not a valid integer. Under set -euo pipefail, this kills the entrypoint before CMD runs. The container exits non-zero, and the test script (also under set -euo pipefail) aborts at the CFG=$(run_override ...) line without printing PASS/FAIL.

Expected Behavior

Invalid overrides should be rejected gracefully — log a security warning to stderr and skip the override, but continue starting the container. The config should remain unchanged (build-time defaults preserved).

Actual Behavior

Container exits immediately with code 1. Test cannot read config. Test script aborts.

Affected

• Nightly runtime-overrides-e2e job on main (run 25118534977)
• Sparky runtime-overrides dispatch (run 25118528500)

Fix

Change validation failures in apply_model_override() from return 1 to return 0 for non-security-critical rejections (malformed integers, invalid boolean, invalid API type). The security message is still logged to stderr, but the container starts normally with the config unchanged.

Symlink and control-character rejections should remain return 1 (those indicate tampering).

Related

• Fixes regression from fix(entrypoint): relax config permissions before write after CAP_DAC_OVERRIDE drop #2659
• Original issue: bug(e2e): runtime-overrides-e2e crashes immediately at baseline config capture #2653