Skip auto-inserting Rack::Lint on Rack 3+

[dduugg]

Problem

Pitchfork's default app builder injects Rack::Lint into the middleware stack whenever RACK_ENV == "development":

https://github.com/Shopify/pitchfork/blob/v0.18.2/lib/pitchfork.rb#L100

Combined with exe/pitchfork:7's ENV["RACK_ENV"] ||= "development", this means every Pitchfork process that didn't explicitly set RACK_ENV ends up with Rack::Lint in its stack.

On Rack 3, Rack::Lint::Wrapper::InputWrapper no longer exposes rewind (it was dropped from the strict spec). Rails middleware and most third-party Rack middleware still call request.body.rewind unconditionally, so any Rails dev process that switches from Puma to Pitchfork starts 500ing on every cookie- or body-bearing request:

undefined method 'rewind' for an instance of Rack::Lint::Wrapper::InputWrapper

This is #177 — the linked issue documents one observable failure (request.raw_post returns nil via Rack::MethodOverride exhausting the stream, then ActionDispatch::Request#reset_stream can't rewind). We hit it in a different shape: ActionDispatch::Cookies and third-party analytics middleware call rewind directly without respond_to?(:rewind) guards, so the failure happens before any controller runs, and the entire Rails GraphQL surface goes dark in dev.

Fix

Gate the auto-insert on Rack::RELEASE < "3". Rack 2 users keep the existing dev affordance. Rack 3 users stop tripping over the incompatibility.

The integration test uses the existing test/integration/fails-rack-lint.ru fixture (which returns a response header that Rack::Lint would 500 on). With lint: true (the spawn_server default — RACK_ENV stays at the CLI's "development" default), the test asserts the response comes through unchanged on Rack 3+, which is only true if the gate kept Rack::Lint out.

Alternatives considered

Issue #177 surfaced three other options, none of which I picked:

• Drop ENV["RACK_ENV"] ||= "development" in exe/pitchfork. Cleanest from a Rails-convention standpoint (RAILS_ENV is the canonical source), but it's a behavior change for every Pitchfork user, not just Rack 3+ users, and it's load-bearing for the dev affordances on Rack 2.
• Replace Rack::Lint with a custom validator that delegates rewind to the underlying input. More invasive, more code to maintain, and it leaks lint-spec violations of the rest of the spec we don't want to swallow.
• Force users to set RACK_ENV=production (or equivalent) themselves. This is the workaround the issue lands on, but it's surprising — the trap fires the moment someone runs bin/pitchfork without reading the issue thread first.

The Rack-version gate is the minimum change that removes the trap for the affected users without changing anything for the unaffected ones. Happy to take it in a different direction if you'd prefer one of the alternatives.

Notes

• Existing tests that explicitly opt out via lint: false (which sets RACK_ENV=production in the spawned process) are unaffected.
• The single existing test that depends on Lint's strict behavior (test/integration/fails-rack-lint.ru is referenced only as a fixture; no test references it yet) is now covered by the new test — but only for the "Rack::Lint is not there" direction, which is what this PR introduces.