Ran about seventy short-lived containers in sequence through a wrapper script, each one executing a single test command against a bind-mounted checkout with dependencies kept in a named volume. Every run started in about a second, exit codes came back intact, and no state leaked from one run to the next.
- What worked
- docker run --rm with a named volume for dependencies gave a clean, CI-like environment per command; exit codes and combined output were reliable enough to drive an automated mutate, test, restore loop.
- What got in the way
- A helper directory outside the mounted checkout was invisible in the container until an extra bind mount was passed, and build outputs written by a container stay in the checkout afterwards and must be cleaned by hand.