Collapse per-kind backend lookup to one call

[Firehed]

Slice S3.8d — Plan 0002 Step 3b, §5.6; RFC 1 §5.1, §8.1.

SymbolBackend carried lookupClassLike and lookupFunction separately, so the same twenty lines of cache-then-build appeared twice in FilesystemBackend and BuiltinBackend, and OpenDocumentBackend held two parallel array pairs. Adding constants (S3.8b) would have tripled that. The backends now take one lookup(QualifiedName, NameKind): ?SymbolInfo.

SymbolSource is unchanged: it keeps a typed method per kind because §5.1 requires a concrete return type. The composite narrows the marker back once per kind, at one site — the O(kinds) narrowing §5.6 trades against O(kinds × backends) methods.

Kind dispatch lands in two factories, one per metadata route, rather than inline in a backend: DeclarationSymbolInfoFactory (parsed declarations) and ReflectionSymbolInfoFactory (the loaded runtime). §4.5 wants kind dispatch confined to a factory, and its static rule arrives in S4.1 — this puts the code in the shape that rule will expect. Moving reflection into the second factory relocates the Reflection* allowlist entry off BuiltinBackend rather than widening it.

Behaviour-preserving: every Step P golden is unchanged, and no golden was recaptured. The phpstan-baseline.neon diff is message text only, following the allowlist message edit — count unchanged at 43.

§8.1 mechanism

The slice also carries §8.1's mechanism for §5.1: SymbolCoverageGridTest, a backend × kind × query grid. Both axes are derived — rows from CompositeSymbolSource::$backends, columns from NameKind::cases() × GridQuery::cases() — so a new kind or backend adds cells that did not exist when the file was written. Every cell either answers over the fixtures or is registered against a named blocker; an unregistered cell fails, and a registration on a cell that now answers fails too, so a blocker cannot outlive its gap.

The grid found one gap on its first run, filed as manifest row SC.16: SymbolExtractor emits no SymbolKind::Constant, so an open document's global constants are never indexed and OpenDocumentBackend::childrenOf cannot enumerate them — while both on-disk and built-in enumeration report constants. WorkspaceNamespaceSource already maps the kind; only a hand-built index in a unit test ever reached that arm. Not fixed here (out of scope); the cell is registered against SC.16.

Acceptance criteria

• SymbolBackend carries one kind-parameterized lookup, not a method per kind
• SymbolSource keeps its typed per-kind methods (§5.1 concrete return types)
• SymbolInfo marker on ClassInfo and FunctionInfo; Formattable not reused
• Kind dispatch confined to factories, not backends
• §5.1 grid with derived axes; unregistered cell fails; stale registration fails
• Every not-applicable cell names a slice id or RFC section
• Behaviour preserved — Step P goldens frozen, none recaptured
• Baseline totals unchanged (43 → 43); no violation absorbed
• 100% line and method coverage on every changed and new class

Sets up S3.8b, whose acceptance is that its diff touches no SymbolBackend implementation.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.62%. Comparing base (85c47fb) to head (04d0f1d).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@             Coverage Diff              @@
##               main     #413      +/-   ##
============================================
- Coverage     98.62%   98.62%   -0.01%     
- Complexity     1810     1814       +4     
============================================
  Files           125      128       +3     
  Lines          4594     4571      -23     
============================================
- Hits           4531     4508      -23     
  Misses           63       63              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.