Run bootstrapFiles per forked worker instead of inheriting the parent's

[ondrejmirtes]

Fixes the private-packagist crash after fork-parallel arrived (Internal error: Failed to start transaction, driver exceptions surfacing string SQLSTATE codes).

Mechanism, verified from the CI stacks and the project's config: their phpstan-dba bootstrap file boots the Symfony kernel and pins the entity manager's live PDO ($em->getConnection()->getNativeConnection()); phpstan-doctrine's objectManagerLoader connects the same way. bootstrapFiles ran once in the parent, so every forked worker inherited the one open postgres socket — and phpstan-dba executes the analysed SQL inside transactions during analysis, so N workers raced the wire protocol on a single shared fd. Spawned workers never had the problem because each re-ran the bootstrap in its own process.

Fix: the analyse flow defers bootstrapFiles and runs them at exactly the places that need them — every BootstrapFilesRunner::run() call site is a place where the files are known to be needed:

• the main thread right before an in-process analysis (AnalyserRunner, at the parallel/in-process routing decision — also covering its fall-throughs: no main script, zero-process schedule);
• every worker, spawned or forked (WorkerRunner — WorkerCommand defers too, so both worker kinds bootstrap identically in their own process);
• the main thread after the workers of a parallel analysis (AnalyserRunner, once the loop is over and no more workers fork), and after a fully cached run (AnalyseApplication's zero-files path) — before the phases that may reflect analysed code (stub validation, collector rules).

Every other command keeps the eager begin()-time execution. The once-per-process latch remains for one documented case: a worker forked from a parent that already ran the files (the fixer flow runs them eagerly before its repeated analysis rounds) inherits that execution.

Autoloaders the files register merge into the source-locator globals at the deferred point — the locators read them lazily. A bootstrap failure in a forked child exits the worker with the error already printed to its collected stdout instead of unwinding past the fork point.

Verified with a bootstrap file that logs getmypid() and defines a constant the analysed code uses: fork = 6 worker PIDs + the parent after them; spawn = 6 worker PIDs + the parent after them; --debug = the parent only, before analysis; fully cached rerun = the parent only. All analyse clean; full test suite, make phpstan (baseline untouched), and CS green.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GnwgpaeUXRkgSDyg95tfK8

[staabm]

Maybe its worth adding private-packagist to our CI pipeline

[ondrejmirtes]

It's in there, but it's 🔮 private 🔮