Implementation self-check

Input: assigned requirements, scope, implementation and repository verification conventions. Return checks, results, uncovered behavior and reproduction details. Use only the caller's authorized files and execution tools.

1. Map each assigned requirement to behavior the user or consumer can observe. Separate general quality gates from checks that establish a specific criterion.
2. For a bug, reproduce it before you change any code, on the surface where it was reported, and record what you observed: that observation is the baseline. Write the regression test and see it fail for the pre-fix behavior, then fix, then run the same reproduction again. Report both observations. A fix written before the defect was seen is a hypothesis: say so, and say which part of the claim remains unverified when the reproduction or the base could not be run.
3. Add or update tests for changed behavior, including relevant boundaries and error states. Every case a criterion lists (states, variants, inputs) gets its own test, or is reported uncovered with the reason. Check that they can fail for the defect: force each condition the change adds or alters to always hold, then never hold, remove each value it adds that a criterion depends on (a class, a key, a label), and see a test fail each time. When you add or change a test, apply tests in the diff.
4. Examine the changed boundary's consumers and real edge cases. For UI changes include semantics, keyboard access, focus, and existing selector contracts; for asynchronous work consider cancellation, retries, ordering and cleanup where affected.
5. Execute the repository's checks and observe relevant runtime behavior. The separate collect-evidence skill can supply mechanics after you select the checks; it does not decide coverage for you. If unavailable, use documented project commands and report the limitation.
6. Report what passed, failed, was not executed or is non-conclusive. Identify remaining criteria and setup recipes. Never label your self-check an independent review.

If repeated attempts fail, stop patching on the same unproven premise. Isolate a smaller reproduction, compare competing causes and escalate a plan or product decision that the evidence disproves.