Defer instructionsToState evaluation in toTest (-62% wall time on a 12-worker run)

[CharlonTank]

Summary

toTest evaluates instructionsToState instructions eagerly while constructing the Test tree, only to read state.testName. Under elm-test's parallel runner, every worker process forces the whole test tree on startup, so each of the N workers re-runs the full E2E simulation for every test — even though each worker only reports results for its 1/N partition.

This PR defers the simulation to inside the Test.test lambda so that each worker only simulates the tests it actually runs. The test name is read via a new private extractTestName helper that walks the EndToEndTest chain without applying any NextStep / AndThen functions (cheap pattern unwrapping only).

Diff (essentials)

extractTestName : EndToEndTest ... -> String
extractTestName instructions =
    case instructions of
        Start state -> state.testName
        NextStep _ inner -> extractTestName inner
        AndThen _ inner -> extractTestName inner

toTest : EndToEndTest ... -> Test
toTest instructions =
    Test.test (extractTestName instructions)
        (\() ->
            let state = instructionsToState instructions   -- moved inside the lambda
            in case state.testErrors of ...
        )

Impact

Measured on a real-world Lamdera app with 14 E2E tests + 56 unit tests, 12 elm-test workers, cold runs on macOS arm64:

	Wall (median of 3)	Duration	Passed
Before	83.0 s	62.0 s	70/70
After	31.2 s	9.4 s	70/70

That is −62 % wall time, −85 % cumulative test duration, with all tests still passing.

The before-numbers also explain previously-mysterious CPU figures: cumulative user CPU was ~580 s for an 83 s wall, i.e. each of the 12 workers was burning ~50 s of CPU re-doing simulation work it would never report. After the patch, workers only simulate their assigned tests.

Test plan

• elm-format src/Effect/Test.elm (no diff)
• lamdera make succeeds with no warnings
• elm-review passes (no new errors; 4 pre-existing suppressions untouched)
• Run the existing test suite: elm-test --compiler=lamdera — same passes/failures as before, faster wall time

Backwards compatibility

No public API change. The EndToEndTest constructors are not exposed from the module, the new extractTestName is private, and toTest's type signature is unchanged. Test names still display identically — they're just read via constructor unwrapping rather than running the simulation.

Notes

The test name walk reads the testName from the original Start state set by start. There is currently no public API in Effect.Test that renames a test mid-chain via a NextStep, so this is always the correct name. If a future API ever lets users rename tests via state updates, extractTestName would need to apply such updates — but that case doesn't exist today.