Testing

cargo test                       # all unit tests: no cloud or Docker needed
cargo test vm::launch            # one module (substring match)
cargo test quota_blocks_every_type_in_the_region   # one test
cargo clippy --all-targets -- -D warnings
cargo fmt --check
cargo test --test vendor_stubs   # every cloud adapter against a stateful stub of its API
cargo test --test docker_e2e -- --ignored --nocapture   # end-to-end, needs Docker

Layers

Layer	How it's tested
Executor (machine/)	a toy machine: journaling, retry-budget exhaustion, resume from a journaled state, cancel → compensation, cancel → resumable
Launch / teardown / k3s machines	provider::fake::FakeProvider, a small simulated cloud with scripted per-offer behaviors (Boots, CreateFails, CreatedButErrors, BootFails, NeverBoots, …), injectable delete failures, a call log, and a max-concurrent-instances counter for leak and duplicate checks
Adapters	pure parsing and classification tests, plus wiremock HTTP servers for each HTTP cloud
Adapters + state machines	vendor API stubs: stateful simulations of each vendor's API (moto for AWS), five shared scenarios per vendor
End to end	tests/docker_e2e.rs: the Docker cloud plus real SSH. It provisions, runs a command, re-provisions (reuse), simulates a crash after create (adopts the container, no duplicate), and tears down.

Async tests run on paused time (#[tokio::test(start_paused = true)]), so the machines' 5-second polls and 30-second backoffs cost nothing. Database tests use Db::open_in_memory().

The launch-machine suite pins the behaviors that matter most:

• capacity triggers compensation and moves to the next region; quota skips the whole region; unavailable skips without teardown
• a transient create error after the request landed adopts the instance instead of creating a second one
• a fatal error after create tears down before aborting
• a failed teardown becomes LEAKED, and the next run cleans up first
• a crash while Booting resumes without a second create
• an unsupported feature is rejected even when a running VM could be reused, and the VM's record is left untouched

CI

Workflow	Runs
.github/workflows/unit.yml	fmt, clippy (-D warnings), unit tests on Linux and macOS, MSRV (1.91.1) check
.github/workflows/integration.yml	vendor-stubs (every HTTP cloud adapter against its stub), aws-moto (AWS against the moto emulator, as a service container), and docker-e2e (the Docker e2e suite plus the acceptance-harness self-test, with GPUPROV_E2E_REQUIRE_DOCKER=1 so a missing daemon fails the job instead of skipping)
.github/workflows/docs.yml	strict MkDocs build on docs PRs; deploys to GitHub Pages on pushes to main

Manual testing against real clouds

Use GPUPROV_HOME=/tmp/gp to keep state away from ~/.gpu-provisioner, and RUST_LOG=debug for verbose progress. gpuprov status <name> shows the journaled transitions with failure details.