feat(objects): a keyed array literal is a valid default for a nullable-array property

[Guikingone]

Fixes #688.

Declaring the class was enough to be refused -- no read, no write, no iteration:

class C { public ?array $x = ["k" => 1]; }
EIR backend error: unsupported EIR backend feature:
object_new for default value of property $x with PHP type Union([Array(Mixed), Void])

What was left

Most of #688 had already been fixed: a mixed/union slot learned to box an INDEXED
array literal, which covers the issue's three reported rows -- ?array $x = [1, 2],
?array $x = [], and mixed $x = [1, 2]. The KEYED spelling had no default form at
all and still failed with the same message.

PHP has no separate associative array type, so ["k" => 1] in a ?array slot is the
same default as [1, 2] is; only the storage a string key needs differs.
LiteralDefaultValue::BoxedAssocArray is the keyed sibling of the existing
BoxedArray, built from the classifier's existing AssocArray arm and the boxing
step BoxedArray already performs.

Both emitters use the OWNED boxer, for the reason recorded on BoxedArray: the
literal allocated the container and the box takes its own reference, so the plain
boxer retains without releasing and leaks one block per object.

Verified

Eleven shapes against host PHP 8.5.10, all byte-identical: the keyed and positional
spellings on ?array, the empty literal, mixed holding a keyed and a positional
literal, array|string, a static nullable-array property, and -- beside them -- the
shapes that already worked, so a future change cannot fix one and lose the other: a
plain array slot, a mixed slot holding a scalar, an explicit null default, and
the constructor-assignment workaround the issue documented.

The classifier arm was confirmed load-bearing by disabling it: both tests fail with
the original object_new for default value of property $x with PHP type Union([Array(Mixed), Void]).

--gc-stats reports allocs=2000 frees=2000 over 200 iterations, and a heap-debug
test pins a clean summary across the keyed, mixed-keyed and positional defaults
together.

cargo test --test codegen_tests -- codegen::objects passes in full (256), as do the
1680 lib tests and 1534 error tests.

One shape is still refused, and it is not this one

A default whose ELEMENTS are array literals -- public array $x = [[1], [2]]; -- is
refused, and a plain array property fails exactly as a ?array one does, so it is
not about the slot at all. LiteralArrayElement has no nested-container form. Filed
as #1052; the same literal is accepted as a local, a parameter default and a class
constant.

Docs and example

docs/php/classes.md states that nullable and union array properties take either
literal spelling, and states the nested-element limitation with its workaround rather
than leaving it to be discovered.

examples/typed-properties/main.php gains a Request whose ?array defaults use
both spellings beside a mixed one and a null one, and clears one to null to show
the slot still holds it. The example is pinned by
test_example_typed_properties_compiles_and_runs, whose expectation is updated in the
same commit and is verbatim host PHP 8.5.10.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KSAAWPyNBq6dP2b5puN3wr

[greptile-apps]

Confidence Score: 5/5

The PR appears safe to merge; the implementation consistently handles keyed defaults across instance and static paths and includes focused ownership coverage.

Summary

This PR adds support for keyed associative-array literal defaults on nullable, mixed, and other union-typed properties.

• Introduces a boxed associative-array default representation and emits it for instance and static properties.
• Uses owned boxing to preserve correct reference-count ownership.
• Adds behavioral, example, and heap-debug regression coverage.
• Documents supported array-default forms and the remaining nested-array limitation.
• Moves the ownership regression into the required runtime GC test area, fully addressing the previous review finding.

_{Reviews (4) · Last reviewed commit: "test(runtime_gc): file the boxed-default..."}

[Guikingone]

Rebased onto current main; this branch is up to date and mergeable again.

Heads-up on an overlap found while rebasing. This PR and #1069 both add the same
LiteralDefaultValue::BoxedAssocArray { value_type, entries } variant to
src/codegen/literal_defaults.rs, arrived at independently:

• here, because public ?array $x = ["k" => 1]; had no boxed form for a KEYED literal (An array literal default on a ?array or mixed property fails to compile #688);
• in feat(objects): a property default may hold array literals, at any depth #1069, because the nested-literal work needed the same keyed boxed form to finish the
  ?array $x = ["k" => [1]]; row of A property default whose elements are array literals is refused #1052.

They are the same addition, so whichever merges first, the other will conflict on that variant
and on its two match arms (block_emit.rs, objects/property_defaults.rs). The resolution is
to keep one copy of the variant, not both — I will rebase the second one onto the first once
the order is decided.