Test conventions (written once, used by every change)

Layers

Layer	Where	Naming
Acceptance (one per scenario)	IronPipeline.Sim.Tests/<Area>/<Name>AcceptanceTests.cs	S<NN>_<Scenario>
Property (invariants)	IronPipeline.Sim.Tests/<Area>/<Name>PropertyTests.cs	<Invariant> e.g. SupplyIsConserved
Golden replay	tests/replays/<name>.replay + hash	run by IronPipeline.Tools replay
Scene/UI wiring	IronPipeline.Game/tests/ (GdUnit4, C#)	<Scene>_<Behavior>

Seed state, never play to it

Build the exact state a scenario needs directly. Until spec 0004 introduces GameState and the builders (GameStateBuilder, UnitBuilder, MapBuilder under IronPipeline.Sim.Tests/Fixtures/), construct value types and maps inline; once they exist, Givens map 1:1 to builder calls. A test never issues commands just to reach its starting position.

Randomness

Once spec 0006 introduces IRandom, tests inject new SeededRandom(1); a scenario that depends on a roll states the seed in its Given. Never use System.Random in tests.

Oracles are numbers

An acceptance test asserts concrete expected values — a number from the Spec, an exact coordinate, an exact count, identity (a == b) — never a relative claim like "greater than before" or "not null" alone. Property tests assert invariants over generated inputs (FsCheck), bounded to realistic ranges.

Golden replays

A replay is seed + command log → final state hash. Changing a hash on purpose requires a one-line reason in the PR ("Decisions made at the build gate") and the save-and-replay-compat reviewer's sign-off. Never regenerate hashes to make a test pass.

Scene tests (GdUnit4)

• Construct state, hand it to the scene, assert on emitted Commands/events and Sim queries — never on Godot node trees, class names or pixel positions.
• Await signals, never Task.Delay / timers.
• One scene-test file per Features number so parallel work never edits the same file.
• Feel is not tested here; that's the playtest checklist.

Refactor evidence

Every batch (behavior → test → green) leaves evidence of the refactor step: a refactor: commit, or a no refactor: <reason> line in the batch's commit body. Tests are frozen during the refactor. Reviewers check "refactor evidence present per batch".

Smells that fail review

Sleeps · commanding through setup · asserting on engine internals · an acceptance test with only relative or null-check assertions · tests edited during a refactor step · a [Fact] that duplicates a property · scenario IDs that don't match the Features file · a batch with no refactor evidence.