Key the dummy app on every input that builds it

[Fivell]

Three defects in how the dummy app is built and reused. All three hide failures rather than cause them, which is why none showed up as a red leg.

The cache key ignored the Active Admin version

The app was generated into spec/rails/rails-<rails version> and reused whenever that directory existed. But active_admin:install and formtastic:install run inside that app, so the Active Admin version is an input to what gets generated.

So the README's own documented invocation was a lie:

RAILS=7.1.0 AA=3.2.0 bundle install
RAILS=7.1.0 AA=3.2.0 bundle exec rspec spec   # reuses the 3.5 app

The leg reported on an app it never built. CI never noticed because runners start empty, so every cell generated its own.

Now keyed on both: rails-7.1.6-aa3.2.0. Verified by switching AA between runs and watching it regenerate.

A half-built app was cached forever

system 'rake setup' unless File.exist?(ENV['RAILS_ROOT'])

Two problems in one line. rails new creates the target directory before the template runs, so any template failure leaves the directory in place — and every later run then skips regeneration and dies on config/environment.rb with a LoadError that points nowhere near the actual cause. And system's return value was discarded, so a failed build proceeded as if it had worked.

This is not hypothetical: a stale half-built app was sitting in this working tree when the harness was first written, and it took a while to work out why.

Now it checks for config/environment.rb — the file the suite actually loads — clears the directory first, and aborts if the build fails. tasks/test.rake likewise aborts rather than ignoring rails new's exit status, and prints the versions it is building for.

spec_helper overwrote the variable the Gemfile reads

ENV['RAILS'] = Rails.version

CI sets RAILS=7.1.0; the Gemfile asks for ~> 7.1.0 and resolves 7.1.6. This line then rewrote ENV['RAILS'] to 7.1.6, and the rake setup child inherited it — so the child's Gemfile said ~> 7.1.6 while the lockfile recorded ~> 7.1.0. Bundler sees a changed dependency and re-resolves, rewriting the lockfile as a side effect of running tests. Under BUNDLE_FROZEN it would abort outright.

The round-trip bought nothing — the only consumer was the path on the next line. Gone; the path now comes from TestAppPaths, shared by the spec helper and the rake task.

The naming approach is lifted from active_admin_import, which already keys its dummy apps this way.

19 examples, 0 failures, before and after.