Preserve diagnosis as read-only unless the user also asks for a fix.
- State the exact observable symptom and establish a fast deterministic reproduction or the closest controlled probe available.
- Minimize the reproduction and rank falsifiable causes from available evidence. Test one cause at a time.
- Trace the failing path through its real public seam and affected external boundary. Add temporary instrumentation only where it distinguishes causes.
- When a fix is authorized, first preserve the symptom as a red-capable regression test at the cheapest layer that proves it completely.
- Apply the smallest causal correction, rerun the focused test and original scenario, then remove temporary instrumentation.
For performance regressions, measure a comparable baseline before changing code and inspect unbounded queries, fan-out, concurrency, memory, collections, and per-item external calls before micro-optimizing.
Report the actual cause, the evidence that falsified alternatives, the changed behavior when fixed, and any remaining uncertainty. Do not convert an unknown or corrupted state into success with a catch, retry, fallback, or timeout increase.