<!-- Generated from models/_canonical by scripts/build-model-variants.js. Edit the canonical source, not this file. Behavioural profile for DeepSeek: scripts/model-profiles.json -->

#Task boundary

1. Implement exactly the task as stated. Do not add abstractions, options, config, or files the task did not name.
2. Comments, identifiers, commit messages and log strings are English only.
3. Stop when the checklist at the end passes. Do not refactor or "improve" surrounding code.
4. Every checklist item below is backed by an assertion in a test or by pasted command output, never by a sentence.

---

#Purpose

Rules for regression testing: ensuring a fixed bug stays fixed and existing behaviour survives change.

The single rule that matters most: write the failing test before the fix. A test written afterwards proves the code does what it currently does. A test written first proves it catches the bug — because you watched it fail.

---

#The workflow

markdown
1. Reproduce      — a test that fails for the reported reason
2. Confirm        — run it; watch it fail with the right error
3. Fix            — change the code
4. Confirm        — the same test passes, nothing else broke
5. Keep           — commit the test with the fix, in the same change

Step 2 is the one people skip. A test that passes before the fix was never testing the bug.

js
// Named for the behaviour, with the issue referenced for context — not
// "test bug 4471", which tells a future reader nothing.
test("refund of a partially captured payment returns only the captured amount", async () => {
  // Regression: #4471 — refunded the authorised total, over-refunding by the
  // uncaptured remainder.
  const payment = await createPayment({ authorised: 10_000, captured: 4_000 });

  const refund = await refundPayment(payment.id);

  expect(refund.amount).toBe(4_000);
});

Reference the issue in a comment, not the test name. The name must describe the behaviour so a failure is legible without opening the tracker.

---

#What to keep as a regression test

Not every bug needs a permanent test. Keep it when:

1. The bug reached production
2. It involved money, data loss, security or privacy
3. It was subtle — an off-by-one, a timezone, a race, a rounding rule
4. It has recurred before
5. The fix is in code that changes often

Skip a permanent test when the bug was a typo caught in review, or the fix removes the possibility structurally — a type change or a database constraint is a stronger guarantee than any test.

---

Useful helpers when reproducing: test.only to isolate the case while you work (never committed), test.each when the bug is one row in a table of inputs, vi.setSystemTime for date-dependent bugs, and --runInBand / --pool=forks when a bug only appears under parallel execution.

For a race, Promise.all over the same operation is usually the shortest reproduction:

js
test("concurrent claims cannot double-spend a credit", async () => {
  const credit = await createCredit({ amount: 100 });

  const results = await Promise.allSettled([
    claimCredit(credit.id), claimCredit(credit.id), claimCredit(credit.id),
  ]);

  expect(results.filter((r) => r.status === "fulfilled")).toHaveLength(1);
});

#Choosing the level

Put the regression test at the lowest level that reproduces it:

Concretely: a rounding bug is expect(total).toBe(4000) in a unit test; a tenant-leak bug is a findFirst assertion in an integration test; a silently broken button is a getByRole assertion in an E2E test.

An E2E test for a rounding bug costs a thousand times the runtime and is flakier. The exception is worth stating: if the bug was the wiring between components, it must be tested at the level where those components meet — that is exactly the seam a unit test cannot see.

---

#Snapshot tests

Snapshots are regression tests that are cheap to create and easy to misuse.

js
// Fragile — any markup change churns it; reviewers approve blindly
expect(render(<Invoice {...props} />)).toMatchSnapshot();

// Focused — asserts the thing that must not regress
expect(screen.getByTestId("total")).toHaveTextContent("£40.00");

1. Keep snapshots small and targeted. A 600-line snapshot is never reviewed; it is regenerated.
2. Never run --update-snapshots to make CI green without reading the diff. That is the mechanism by which a real regression gets committed as expected behaviour.
3. Prefer explicit assertions for anything with business meaning. Reserve snapshots for shapes with no better expression.

---

#Guarding against silent regressions

Some regressions are better prevented structurally than tested:

ts
// The compiler now fails when a new status is added and left unhandled —
// stronger and cheaper than a test asserting the same thing.
function label(status: Status): string {
  switch (status) {
    case "draft": return "Draft";
    case "sent": return "Sent";
    default: {
      const _exhaustive: never = status;
      return _exhaustive;
    }
  }
}

#Keeping the suite honest

A regression suite accumulates. Without maintenance it becomes archaeology — tests nobody understands, guarding behaviour nobody wants.

1. Run the full suite before shipping a fix — npm test locally, not just the one file. A fix that repairs refundPayment and breaks capturePayment is caught only by the rest of the suite.
2. Delete tests for removed features. A test for deleted code is pure cost.
3. Consolidate when six tests cover one rule through slightly different paths.
4. Fix or quarantine flaky tests immediately. One test retried until green teaches the team to ignore red, which is how a real failure ships.
5. Never comment out or .skip a failing test to unblock a release without an issue and an owner. A skipped test is deleted coverage that still looks present.
6. Re-read the suite when a module is rewritten. Tests asserting the old design block the new one for no benefit.

---

#Anti-patterns

---

#Checklist

• Every bug fix ships with a test in the same change
• The test was observed failing before the fix was applied
• Test names describe behaviour; issue references live in comments
• Each regression test sits at the lowest level that reproduces the bug
• Wiring bugs are tested at the seam, not below it
• Snapshots are small, targeted and diff-reviewed
• Snapshot updates are never applied blindly to make CI pass
• Flaky tests are fixed or quarantined with an owner, never skipped silently
• Tests for removed features are deleted
• Structural fixes — types, constraints — are preferred over tests where possible