dev-verification

Run the original symptom reproduction or regression test fresh and read the full output including the exit status. Code changed is not bug-fixed evidence; only a passing check that exercises the original failure proves the fix.

Run the focused or full test command and read the complete output showing zero failures, not a piped tail or a wrapper's exit code. A multi-case test stops at its first failure, so fix the first failure and re-run before counting coverage.

A test asserting the code's own output proves stability, not correctness. Certify correctness by checking against an independently derived reference that shares no code path, such as a textbook formula, autodiff comparison, or brute-force enumeration.

Mark a finding repaired and pending review when the fix is written, but only clear it after a pass that did not write the fix re-reads the code. Re-review the code itself, not just the finding list, since repairs can introduce new defects.

Compare accumulated CPU time from ps -o time against wall-clock elapsed time. Near-zero CPU growth over minutes indicates the job is blocked or hung rather than busy, and resource-shaped parallel-test failures should be rerun with lower parallelism first.