Bucket the last 52 flat-namespace classes into src/

[mastacontrola]

Pages, hooks, reports and events were the only core classes left in a flat
namespace FOG;, sitting under lib/ with a class_alias trailer each. They are
PSR-4 files under src/{Pages,Hooks,Reports,Events} now, declaring
FOG\Pages\HostManagement, FOG\Hooks\BootItem, FOG\Reports\Audit_Report and
FOG\Events\HostList, with the 52 aliases deleted. There is no flat FOG\
namespace in core any more, and every lib/ subdirectory except router/ and
plugins/ is empty of code.

Why ADR 0013 said this was impossible, and why it wasn't

ADR 0013 recorded that these could not move because "PSR-4 does not do discovery".
That was true and was not the constraint. The constraint was that the three
discovery sites derived a bare class name from basename($file) and resolved it
from the global namespace, so the files needed an alias — and the alias was exactly
what the rest of the migration was retiring. Discovery had to learn a second file
shape first; nobody had taught it one. That is a property of the loader, not of
PSR-4. The ADR carries the correction.

How discovery works now

Each site reads two sources and merges them: FOGBase::coreitems($bucket) for core,
and the existing fileitems() walk for plugins, which keep the
<plugin>/<dir>/<name>.<type>.php shape and the global namespace (ADR 0009).

coreitems() filters Initiator::srcFileList(), which is already built and already
cached, so it adds no directory walk to a request. FOGBase::classFromDiscoveredFile()
then derives the name — stripping the discovery extension when the file carries one
and .php otherwise — and hands it to qualify(), which spans both shapes with no
branch on provenance: it maps a name src/ declares onto its FQCN and passes anything
else through untouched.

startClassFromFiles() takes the extension rather than a negative string length, since
it can no longer assume one shape.

The one thing that was nearly an authorization change

A report's filename used to be lowercase, and three things read it as such: the menu
label, the base64 f URL parameter, and the keys of Authorization::REPORT_NODES —
which is a permission gate. The PSR-4 filename is Audit_Report.php, so
loadCustomReports() now lowercases what it used to get lowercased for free. Without
that one call the same report answers to a different URL and a different permission
node than it did before. That is a silent authorization change, not a cosmetic one,
and tests/usertracking-permission-split.test.php now pins the real transform rather
than a hardcoded list.

Verification

Booted a pre-move and a post-move tree against the live database and diffed what
discovery actually found — report labels, page registration, core hook and event
resolution, and both listener registries. Byte-identical.

The probe was mutation-tested, not assumed: making coreitems() return [] takes
reports 17 → 4 and pages 1 → 0. That same mutation exposed a hole in the first version
of the probe — every core hook ships $active = false, so the listener registry is
identical whether core hook discovery works or finds nothing — and the probe now
resolves those classes directly.

• Test suite: 238 passed, 0 failed, matching the pre-change baseline exactly.
• Both phpstan passes: [OK] No errors.

tests/psr4-bridge.test.php gains a second probe file for the bridge's
case-insensitive prefix. The assertion it replaces, in autoload.test.php, was a fake
gate: it ran after the class had already been loaded, so PHP's own case-insensitive
class table answered and the bridge was never consulted. Mutating strncasecmp to
strncmp did not fail it. The new one does.

tests/stale-class-file-list.test.php had a real bug of the same family — it passed
-strlen('.hook.php') where the parameter is now the extension itself.

Notes for reviewers

• The phpstan baselines were repointed by hand — 190 path: lines and 81 message
  FQCNs — rather than regenerated, which would have swept up unrelated drift. One
  pre-existing stale entry for inventory_report.report.php is deliberately left alone.
• bin/psr4-scan.php buckets the four kinds by ancestry rather than 52 table rows.
  The rule order matters: ReportManagement before FOGPage, since a report's chain
  reaches both.
• The 13 report classes gained use FOG\Pages\ReportManagement;. They extends ReportManagement unqualified, which used to resolve through the bridge.

Downstream — this needs FOGProject/fog-plugins#32

Merge this first, then the plugins PR, then cut a fog-plugins release.

Out-of-tree plugins spelling \FOG\ReportManagement need \FOG\Pages\ReportManagement.
Eight bundled reports in fog-plugins do exactly that, and they fatal the moment someone
opens the report once this lands. The bridge refuses the flat spelling with an error
naming the correct FQCN rather than resolving it, so the failure is legible — but it is
still a failure.

Note that fog-plugins' own tests/core-references-are-qualified.test.php does not
catch this and cannot: it is a syntax gate asking whether a class reference carries a
leading backslash, not whether the name resolves. It stayed green through the whole
breakage.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VTEiBnDpNrgzu1tSqt1318