Test Normal, Edge and Invalid Plus Old Features

Why this matters

The classic beginner arc: fix the validation message, break the echo; fix the echo, break persistence; fix persistence, break the layout. Each repair untested, each test skipped "because the change was small". Small changes break things precisely because they touch shared code. A five-minute routine after every fix ends the arc.

Concepts

Three input classes. Normal (2 loaves, Friday 10:00): the everyday case must pass. Edge (1 loaf minimum, 50 maximum, longest dish name, narrowest width): boundaries must hold. Invalid (empty, spaces, letters in quantity, past time): guards must answer with sentences, values kept. Lesson M05L17 to M05L19 built the guards; this lesson proves them per fix.

Old features re-run. Every fix ends with the regression pass: echo still echoes, validation still messages, list still persists, layout still fits at 360px. Four old checks, two minutes. A fix that repairs one thing by breaking another is not a fix; it is a move. The routine catches moves.

Test list before fix. Write the checks the fix must pass before changing anything (the acceptance criteria habit from M02L8, now per bug). Run the list after. "It opens" never appears: opening proves rendering, hands prove behavior (fill, submit, reload, delete, break on purpose).

Worked example

Fix: storage quota check plus message. Test list: normal submit persists; edge: exactly-50 loaves accepted; invalid: empty still guarded with sentence; old: echo intact, validation messages intact, 360px layout intact, delete works. Run: all pass in four minutes. Without the list, the layout check would have been skipped, and the message styling did overflow at 360px: caught only because the routine is unconditional.

Expected result: a written test list run after your next fix with normal, edge, invalid and old-feature lines all checked off.

The common wrong move

Testing only the reported case ("the reported bug is fixed, ship it"). The reported case is the minority of behavior. Routines test the neighborhood, not just the address.

Lab and next step

Lab A21 runs the full routine on a provided fix with a planted regression. Next, lesson 28 breaks the most expensive loop: asking AI for the same fix repeatedly.