test(magecommand): synthetic di verify fixtures for the omitted-artifact shapes

[pingiun]

Every di verify regression that reached a client pipeline had one shape: an archive from a Magento that compiled every module on disk, checked against magecommand, which compiles only the enabled ones. The extra artifacts are expected, and the classifier's job is to say so.

Three times it said so for only part of a component:

bug	effect
owning module keyed off the first two path segments	matched nothing once paths carried the code/ prefix — 254 files explained in one layout, 0 in the other
only /Interceptor.php recognized	a disabled module's factories, proxies and extensions reported raw
class resolved to its source file	a generated name has none, so plain classes were explained while interceptors and factories were not

None of these needed a real store to catch. They need an archive holding artifacts the output lacks, in those shapes — so the fixture compiles a synthetic root for real, snapshots that output as the archive, and adds what the older compiler would also have written. No PHP, no corpus.

Cases

• A disabled module's artifacts, every generated kind (interceptor, factory, proxy, extension, extension-interface) — asserted in both directory layouts, which is the layout bug exactly.
• An unregistered component's interceptor and factory — files that only resolve through their dispatch subject.
• A missing artifact for an ENABLED module — must still fail.
• Two identical trees — must verify clean.

The last two matter as much as the first two: a rule that explains everything is worthless.

Each positive case fails without its fix

I verified this rather than assuming it — a fixture that passes either way is worse than none:

• removing the code/ prefix strip → the layout case fails;
• narrowing artifact kinds back to interceptors-only → the layout case fails;
• removing the dispatch-subject fallback → the unregistered-component case fails.

Why this suite specifically

di_compile_* fixtures already existed; there was no di_verify_* fixture at all, and every recurring bug lived in the verify classifier. That was the gap.

🤖 Generated with Claude Code