fix(gate-64): a named constant is the prelude, spelled better — not a missing one

[rubenvdlinde]

The finding that had nothing behind it

gate-64 recognised the ADR-040 autoload prelude only as a quoted literal inside the call parentheses:

\OC_App::registerAutoloading('openregister', $path);              // accepted
\OC_App::registerAutoloading(self::OPENREGISTER_APP_ID, $path);   // REJECTED

The second is the same call with the app id given a name, four lines below private const OPENREGISTER_APP_ID = 'openregister';. doriath writes it that way — and gate-64 reported it as an AppHost adoption with no prelude, with the prelude present, correct, and unit-tested.

Measured on doriath#163: FAIL — 2 AppHost adoption(s) with no OpenRegister autoload prelude.

A gate that fails the tidier spelling of the thing it is asking for is not reporting a defect. It teaches people to write the untidy spelling, and it spends the reader's attention on nothing.

The fix, and why it does not go blind

Resolution is deliberately shallow and literal: a name counts only when the same source blob binds it to the exact string 'openregister' — via const, define(), or a plain variable. Imports are not followed, and a name that cannot be seen bound is not accepted.

That last property is the whole design. "Accept any argument" would clear doriath and blind the gate in the same stroke, so two of the four new tests exist to fail in exactly that direction.

test	direction
test_named_constant_bound_to_openregister_is_GREEN	doriath's shape passes
test_variable_and_define_spellings_are_accepted	the other two bindings
test_constant_bound_to_another_app_still_FAILS	registerAutoloading() for a different app is not a prelude
test_unknown_constant_still_FAILS	an unresolvable name is not taken on trust

Mutants — disjoint sets, both directions

mutant	tests turned red
accept any argument name	the 2 discriminators
resolve nothing (regression to the bug)	the 2 acceptance tests

Measured against the real trees

This is what settles it — the gate must lose the false positive and keep both true positives.

repo	gate-64 @ main	with this fix
doriath	rc=1 FAIL	rc=0 PASS	false positive cleared
scholiq	rc=1 FAIL	rc=1 FAIL	real defect, still caught
openbuild	rc=1 FAIL	rc=1 FAIL	real defect, still caught

scholiq and openbuild genuinely have no prelude, and openbuild is the acute one: its class_exists(Bootstrap::class) guard answers false on a perfectly healthy instance, because openbuild sorts before openregister and Coordinator::registerApps() registers autoloading one app at a time. So Bootstrap::register() has apparently never run there, and its generic controllers, install repair steps and deep-link listener are silently absent. Those are fixed in their own repos, not here.

⚠️ Note on my own method: my first before/after table was wrong in the safe-looking direction — I copied the "baseline" helper after patching it, so both columns ran the fixed matcher and doriath appeared to pass on main. The table above is re-measured against git show origin/main: and the md5sums were checked. A baseline captured after the edit is not a baseline.

18 gate-64 tests pass.

🤖 Generated with Claude Code

[rubenvdlinde]

Superseded by #184, which landed while this was in review and is a superset of it.

Both fixes resolve the same false positive — gate-64 matching the app id only as a quoted literal, so doriath's registerAutoloading(self::OPENREGISTER_APP_ID, $path) read as a missing prelude. #184 additionally strips comments before matching, which closes a false GREEN this PR did not address: a commented-out registerAutoloading('openregister', …) would have counted as a prelude that no longer runs. That is the more dangerous direction and it belongs in the gate.

Re-measured against the real trees with #184 as it now stands on main:

repo	gate-64
doriath	rc=0 PASS — false positive cleared
scholiq	rc=1 FAIL — real defect, still caught
openbuild	rc=1 FAIL — real defect, still caught

Identical to this PR's outcome, so nothing is lost by closing it.

The only delta here was accepting two further bindings — define('NAME', 'openregister') and $var = 'openregister' — where #184 accepts const only. No repo in the fleet uses either spelling today, so I am not re-opening it as a follow-up; if one ever does, gate-64 will say so by name and the binding can be added then with its own paired test.

The two real defects this exposed are being fixed in their own repos: ConductionNL/scholiq#289 and ConductionNL/openbuild#139.