test-design-mandates

name	test-design-mandates
description	Use when planning which tests an approved set of examples actually needs before any code is written — where each test enters and what it observes, whether a rule those examples already cover deserves a test of its own, and whether a scenario reads in business terms and ends in an outcome a user could observe.

Test Design Mandates

Overview

Five mandatory rules that govern how tests are designed in a Clean Architecture context. Applied after Gherkin scenarios are written, before any implementation starts.

Core principle: Tests enter through a use case boundary and assert at the next visible boundary. Never test internal classes directly.

---

Mandate 1 — Layer Boundary Enforcement

Rule: Every AC names its use case boundary — the Application layer entry point the test enters through. All tests at the Application layer enter through that use case. Assertions are made via application interfaces (repositories, gateways) or the use case return value. Never instantiate an internal domain class directly in a test that is exercising Application behaviour.

Why: Prevents TBU (Tested But Unwired) defects — production code that works in isolation but is never called through the real composition root.

Application:

• Acceptance test → enters through {UseCaseName} use case / command handler
• Domain unit test (when extracted) → enters through the domain function or policy's public signature
• Infrastructure test → enters through the application interface (repository contract)

TBU detection checklist (run after GREEN):

• Can I delete the production implementation and still have the test pass? → Wired? YES = TBU
• Is the use case wired in the DI container? → Not tested if no integration test exercises the wiring
• Does the acceptance test use an InMemory implementation, not the real one? → Correct for unit pass; add integration test for real wiring

---

Mandate 2 — Business Language Abstraction

Rule: Maintain 3 strictly separated layers of abstraction in test code. No technical vocabulary leaks upward.

Layer	Location	Language	Example
Gherkin	.feature files	Pure business	Given a driver with a clean record
Step methods	Step definition classes	Bridge — calls use case	eligibilityUseCase.handle(command)
Business services / Use cases	Test helpers or application	Technical	new CheckEligibilityCommand(driverId, ...)

Violations:

• ❌ Gherkin step: Given the EligibilityApplicationService is instantiated
• ❌ Step method builds DTOs with raw primitives without a factory method
• ✅ Step method: var command = EligibilityCommandBuilder.forDriverWith(cleanRecord);

Implementation:

• Create {Feature}Builder / {Feature}Mother test factories in step definition layer
• Step methods delegate to builders — never hardcode IDs, dates, or primitives in Gherkin or steps

---

Mandate 3 — User Journey Completeness

Rule: Tests validate complete user journeys, not isolated operations. A scenario must include setup, action, and observable outcome — all three, always.

Structure:

Setup (Given)   → state of the system before the action
Action (When)   → ONE business trigger
Outcome (Then)  → observable business result (not internal state)

Incomplete journey anti-patterns:

• ❌ Only testing Given + When with no assertion → setup test, not a behaviour test
• ❌ Then asserting internal state (e.g., field value on domain object) → not observable by user
• ❌ Missing setup → test depends on implicit shared state (test ordering bug)

Journey completeness check: Ask "could a user observe this outcome in the real system?" If yes, the Then is correctly observable.

---

Mandate 4 — Domain Test Extraction (Gated)

Default rule: Acceptance tests do the work. No domain unit test by default, even when a pure function is extracted by refactoring. A regression that breaks 2+ AC together is the correct business signal — the duplicated red is the proof that the rule traverses multiple user journeys, not a redundancy to eliminate.

Anti-pattern this rule prevents — DOUBLE-COVERAGE: A pure function whose every behavioral branch is already reached by planned acceptance scenarios MUST NOT receive a dedicated domain test. Two suites asserting the same behavior drift together on every rule change and add zero failure-discrimination value.

A domain unit test is REQUIRED if and only if one of the two gates opens:

• Gate (a) — Branch unreachable via AC. A behavioral branch of the function exists but no realistic acceptance scenario can trigger it (defensive case, exhaustive-enum fallback, technical guard). The AC physically CANNOT observe it.
• Gate (b) — Combinatorial economy. Covering the input grid through AC alone would explode the Gherkin scenario count (indicative threshold: > 10–15 scenarios for a single rule). Keep 3–5 representative business AC (happy path + key rejections + critical boundaries) and delegate combinatorial sweep to a parameterized domain test ([Theory], @ParameterizedTest, table-driven).

No gate opens → REJECT extraction. Record in the test plan: M4 negative — saturated by AC.

Process for each candidate pure function P:

1. Enumerate B(P) — distinct behavioral branches of P.
2. Enumerate A(P) — branches reached by planned AC scenarios.
3. If A(P) == B(P) AND combinatorial size ≤ 10–15 → forbidden to extract, log M4 negative — saturated by AC.
4. Otherwise, name which gate opens and record the corresponding Extraction Reason code.

Counter-example (STORY-41 contrefactual): Vehicle.MinimumAge() has B(P) = {21, 16}. Five planned AC reach both outputs, so A(P) == B(P). Combinatorial size = 3 cases, well under threshold. Both gates closed → no domain test. The 5 AC do all the work; if MinimumAge() breaks, all 5 turn red — that is the intended business signal.

---

Mandate 5 — Visual/Positional AC Enforcement

Rule: Any AC expressed in visual, positional, or style terms (pixel position, full-page width, background colour, relative alignment between two elements) is closed ONLY by a Playwright E2E test performing a real measurement (boundingBox(), computed colour, computed layout) against a real browser engine. A jsdom/unit test is NOT a valid closure for this AC type.

Why: jsdom has no layout engine (getBoundingClientRect() → zeros). Unit assertions proxy via DOM order/class presence, not rendered proof. Without this mandate, a unit-TDD-only contributor believes the AC is verified while CSS/mount/z-index regressions go uncaught.

Enforcement:

• Tag scenario @visual (see bdd-methodology).
• Add matrix row: Layer = E2E, Double Type = Real browser (Playwright).
• @visual w/o tests/e2e/ spec → coverage gap, block story (G9 check).

---

Coverage Matrix

Build this matrix for every story before implementation starts. Any row with Layer = Domain MUST carry an Extraction Reason code (see Mandate 4). Without a reason code, the row is not authorized — remove it.

Allowed Extraction Reason codes:

• branch_unreachable_via_AC — Gate (a)
• combinatorial_economy — Gate (b)

Scenario	Use Case Boundary	Layer	Extraction Reason	Double Type	Walking Skeleton	Priority
Happy path — driver eligible	CheckEligibilityUseCase	Application	—	InMemory repository	A	P1
Edge — driver at age limit	CheckEligibilityUseCase	Application	—	InMemory repository	A	P1
Rejection — too many accidents	CheckEligibilityUseCase	Application	—	InMemory repository	A	P2
Combinatorial sweep — premium grid	PricingPolicy.computePremium	Domain	combinatorial_economy	None (pure function)	—	P2
Infrastructure — real persistence	IEligibilityRepository	Infrastructure	—	Real DB (Testcontainers)	—	P3

Priority:

• P1 — happy path, walking skeleton basis
• P2 — business rule coverage, edge cases
• P3 — infrastructure, integration, error paths

---

Walking Skeleton Strategy

A walking skeleton is the thinnest possible slice that exercises the full path from use case to output — enough to prove the wiring works.

4 Strategies

Strategy	When to use	Setup
A — Full InMemory	Feature is purely internal (no external services)	InMemory repository, no real I/O
B — Real local + Fake costly	Feature needs a local DB but avoids expensive external calls	Testcontainers DB + fake/stub for external service
C — Real local	Feature integrates with controllable local infrastructure	Testcontainers for all local dependencies
D — Configurable	Feature must run in both unit and integration mode	Strategy/feature flag selects double type at test startup

Decision Tree

Does the feature write to or read from persistent storage?
├── NO  → Strategy A (full InMemory)
└── YES → Does it call an expensive external service (payment, SMS, AI)?
          ├── YES → Strategy B (real local DB + fake external)
          └── NO  → Is the storage local and controllable?
                    ├── YES → Strategy C (real local with Testcontainers)
                    └── NO  → Strategy D (configurable)

Walking Skeleton Sizing

• 2–5 walking skeletons per feature (one per major flow variant)
• 15–20 focused scenarios total per feature (detailed behaviour coverage)
• Tag walking skeleton scenarios with @smoke for fast validation

---

Layer Assignment Rules

What to test	Test project	Layer	Double type
Use case / command handler	UnitTest	Application	InMemory application interfaces
Domain policy / specification	UnitTest	Domain	None (pure function — call directly)
Repository adapter	IntegrationTest	Infrastructure	Real DB via Testcontainers
API controller / endpoint	IntegrationTest	API	In-process app host (WebApplicationFactory or equivalent)
Architecture boundaries	IntegrationTest	Architecture	Static analysis (NetArchTest, ArchUnit, etc.)
Visual/positional/style assertion (@visual)	E2E (Playwright)	Presentation (E2E)	Real browser engine — boundingBox(), getComputedStyle()

Never:

• Test a domain entity by instantiating it directly in an Application acceptance test
• Use a real database in UnitTest
• Use a mock where an InMemory fake exists (InMemory > mock for repositories)

---

References

• coverage-matrix-template.md — blank coverage matrix template
• layer-assignment-rules.md — detailed rules per layer
• boundary-enforcement-principle.md — use case boundary enforcement and TBU prevention
• walking-skeleton-strategy.md — 4 strategies with decision tree and examples