Core rule
Reduce the failure until one controlled variable explains the difference between working and broken behavior.
Start from the real failing path and build the smallest case that can still reproduce the symptom. Remove unrelated files, plugins, configuration, data, and timing assumptions one at a time. Preserve the real boundary that may be causal: the actual parser, shell, serializer, protocol, or filesystem behavior.
Once minimal, vary one factor at a time and record the outcome. A reproducer that fails for a different reason is not useful.
Keep the final case as a regression fixture when practical; it should make the bug cheap to understand and impossible to reintroduce silently.