Fix intermittent server boot failure caused by working directory race

[soutaro]

Motivation

CI hangs intermittently on Ruby 4.0.6 (first seen in #724 / #725) — jobs time out after 25 minutes with the test suite stuck mid-run. The cause is a working-directory race in RunnerClient.create_client, not setup-ruby or gem updates.

create_client checks File.exist?("bin/rails") relative to the process working directory, and spawns the rails runner server with the inherited cwd. But the working directory is process-global state that other threads mutate: RuboCop's config loader evaluates configs with Dir.chdir(config dir) { ERB.new(...).result } while resolving inherit_gem. When the Ruby LSP RuboCop addon activates concurrently with the Rails addon boot thread, the boot thread can observe the rubocop-shopify gem directory as cwd. The bin/rails check then fails and create_client silently falls back to a NullClient.

The race is long-standing but almost never fired until Ruby 4.0.6, whose thread-scheduler changes (Bug #21685 backport) altered thread interleaving enough to make it fire several times per 100 boots on CI runners. That's why bumping setup-ruby past 1.317.0 (which resolves "4.0" to 4.0.6 instead of 4.0.5) broke CI. Trace from a failing boot captured in the investigation PR (#726, now closed):

[RuboCop] Activating RuboCop LSP addon 1.88.2.
Ruby LSP Rails create_client start (cwd=.../vendor/bundle/ruby/4.0.0/gems/rubocop-shopify-2.18.0)
Ruby LSP Rails create_client: bin/rails NOT found (cwd=.../gems/rubocop-shopify-2.18.0)
[RuboCop] Initialized RuboCop LSP addon 1.88.2.

Note that the hang itself only happens in this repo's test suite: several tests busy-wait with sleep(0.1) while ...NullClient for the real client to boot, and the loop never terminates when boot fails because NullClient is the final state. In real editor usage a boot failure does not hang anything — the add-on falls back to the NullClient, logs a warning, and Rails-dependent features are silently unavailable.

Implementation

• Resolve bin/rails against global_state.workspace_path instead of the cwd, and spawn the server with an explicit chdir: to the workspace path, so neither depends on racy process-global state.
• Fix the tests' unbounded busy-wait as well: add Addon#join_boot_thread so the boot outcome can be awaited deterministically, and make the tests join the boot thread and assert that the client actually booted. A boot failure now fails the test immediately with a clear message instead of hanging the suite until the CI timeout.

[jesse-shopify]

I really like seeing the removal of all the sleep calls. This is a much improved design.