feat: subtract a warm-up boot baseline from Tia dependency edges

[iAmKevinMcKee]

The problem

Tia derives dependency edges from per-test coverage windows, and those windows include everything that runs in setUp(). Frameworks that boot inside every test — Laravel being the canonical case — re-execute the same bootstrap lines in every test's window: service providers, bootstrap/app.php, route registration. Filament makes this dramatic: panel route registration executes getPages()/getRelations() in every registered resource class on every boot (verified by dumping executed lines; Laravel route caching does not avoid it).

The result on a production Laravel 12 + Filament 5 app (~100 resources):

• a pure enum unit test recorded 146 dependency edges, including all ~100 Filament resource classes, every service provider, and the core models;
• a dashboard feature test recorded 1018 edges;
• editing any Filament resource — where most day-to-day work happens — invalidated the entire suite. Tia effectively degraded to "always run everything" for the changes developers actually make.

Any Filament project will hit this; any Laravel project hits a milder version (providers/routes as universal edges).

The fix

pest()->tia()->warmupUsing(Closure $callback) — a callback that runs once per process under the coverage driver, before the first test. Every line it executes forms a warm-up baseline that is subtracted from each test's recorded coverage before edges are derived. A file only becomes a dependency when a test executes lines beyond the baseline — so a feature test that genuinely renders a resource's form keeps that edge, while a unit test that merely booted the framework loses it.

// tests/Pest.php
pest()->tia()->warmupUsing(function (): void {
    $app = require __DIR__.'/../bootstrap/app.php';
    $app->make(Illuminate\Contracts\Console\Kernel::class)->bootstrap();

    // Undo global state the boot mutated before returning — see note below.
});

Measured results (same app as above)

Metric	Before	After
Enum unit test edges	146	6 (its enum, TestCase, the tenant models its setUp genuinely touches)
Dashboard feature test edges	1018	27
Edit an unrelated Filament resource	full suite reruns	0 tests rerun (682-test set replays in ~260ms)
Edit a service with 1 dependent test	full suite reruns	exactly 1 test file reruns
Edit a broadly-used model (Project)	full suite reruns	its 57 genuine dependents rerun, rest replay

Warm-up responsibility (important caveat)

The callback runs in the test process, so it must clean up global state it mutates. For Laravel that includes a subtle one we hit in practice: booting an app loads .env through Dotenv, and the static Illuminate\Support\Env repository then treats those variables as Dotenv-owned — when the real test application boots later, Dotenv overwrites the restored APP_ENV=testing with the .env value, which (among other things) silently disables Filament's fillFormDataForTesting() and produces confusing downstream failures. A working Laravel warm-up restores $_ENV/$_SERVER/putenv snapshots, resets Env::$repository via reflection, resets the container/facades, and unwinds error/exception handlers back to their pre-boot state. The warmupUsing() PHPDoc calls this out; if maintainers prefer, a follow-up could ship a ready-made Laravel warm-up (Pest already has Laravel awareness in WatchDefaults\Laravel) so users get this behavior with zero config — happy to do that in this PR or a follow-up, whichever you prefer.

Implementation notes

• The baseline is collected lazily at the first beginTest() with coverage capture active: start driver → run callback → collect scoped lines → reset for the per-test window. No behavior change when no warm-up is registered.
• Subtraction happens in Recorder::filesWithExecutedLines() per file/line; the existing single-line/max-line heuristic is preserved.
• Piggyback-coverage mode (--coverage + Tia) is unaffected — the warm-up only runs when the Recorder drives pcov/xdebug itself.
• Driver-dependent tests probe that the driver can actually instrument the fixtures (pcov only instruments under pcov.directory) and skip otherwise, so they are safe in any CI configuration.

Tests

• warmupUsing() excludes files fully covered by the baseline.
• Files where a test goes beyond the baseline keep their edge.
• No-warm-up behavior unchanged.
• The callback is never invoked without a coverage driver (link-tracking mode).

composer lint, composer test:type:check pass; composer test:unit passes apart from Backtrace::it gets file name from called file, which fails identically on a clean checkout in this environment (pcov extension loaded) and is unrelated.

🤖 Generated with Claude Code