writing-tests

Write the test, then revert the fix and confirm the test goes red for the bug's signature rather than a setup error, then restore the fix and confirm green. State this red-on-old, green-on-new proof when reporting the change.

A text search passes whether or not the matched line ever executes, breaks on harmless renames, and stays green against broken implementations containing the token. Drive the real code path and assert observable outcomes like files written, exit codes, or argv passed.

Stubs are licensed only when the real dependency genuinely cannot run, such as paid APIs or missing hardware capabilities. Feed the stub a recorded, scrubbed response captured from the real dependency, and document why the real thing could not run.

A test serving only agreeing input proves nothing, since it stays green after the verification is deleted. Drive the refusing direction too and assert what did not happen, such as the binary never reaching PATH.

A parametrize over an empty sequence collects zero cases and reports a clean pass. Assert the input list is non-empty at the single point that reads it, so a broken helper cannot silently turn every dependent test green.

Assert the correct behavior and mark the test with pytest xfail using strict=True and a reason naming where the fix lives. Strict mode turns the eventual fix into an XPASS failure until the marker is removed.