Lifts the "non-empty abstract stack at BB boundary" restriction in
the IR Builder. Previously threw on any BB-end with stack residue
(typical javac shapes: ternary-store, short-circuit &&/||, anything
that leaves a value on the stack across a branch).

Now spills residual abstract-stack entries into synthetic local slots
at branch sources (goto/condgoto/ifPop/implicit-fallthrough); target
BBs reload from those slots at entry. JVM verifier guarantees all
predecessors of a merge point deliver the same stack-shape, so per-
target-PC slot allocation works.

New helpers in Builder: bbEntrySlots[targetPc] map, nextSyntheticSlot
counter, allocOrReuseSlots / spillStackToSlot / spillStackForBoth /
reloadStackFromSlots. Method.maxLocals widened to include synthetic
slots.

Real-library probe: 86.5% → 99.3% IR coverage on commons-lang3
(485-fallback drop from this single change). The originally-estimated
1.5x perf headroom is dwarfed by the actual coverage impact.

Final state docs after the 2026-05-03 work plan landed.

ROADMAP.md:
  - Sub-step 1c-β note updated with the empirical coverage finding
    (95% of remaining fallbacks; perf vs coverage impact distinct)
  - New sub-step entries: 1c-ε (switch terminator), 1c-ζ (rare-opcode
    tail)
  - Work-plan section enumerating items php-java#1–php-java#10 in order, with php-java#10
    (remove dead string-path) explicitly DEFERRED — needs rank-1
    evidence on more than one JAR before subtracting 1500-line
    fallback.

STATUS.md:
  - Real-library probe section: 6-row series table showing
    progression 65.6% → 86.5% → 99.3% → 99.9% → 99.9% → 100.0%.
  - Architectural verdict: design holds against production
    bytecode; sub-step 1c-β was the load-bearing lever.

bench/probe-real-library.md:
  - TL;DR updated to the 6-row series view + final 100.0% reading.

The architectural piece. Before: AOT hardcoded `isStatic: true` for
all methods (Builder.php:283, Compiler.php:1046). After: read
MethodInfo.ACC_STATIC; emit instance methods as `public function`
with $L[0] = $this; emit static methods as `public static function`.

Threading: Compiler::compileFromGenericClass classifies each method,
passes $isStatic to tryBuildIrMethod and compileMethod. Builder::
buildMethod and Compiler::compileMethod accept $isStatic; widen
maxLocals to include the implicit `$this` slot for instance methods;
Lowerer prelude seeds `$L = [$this, $__a0, ...]` for non-static.

<init> emit: previously skipped at compileFromGenericClass:242. Now
emitted as `__construct` (the existing mangleMethod already maps
<init>→__construct). super(...).<init>() calls into JDK abstract
bases (Object/Record/Number/Enum/Throwable/Exception/RuntimeException/
Error) emit as no-op via isAbstractBaseInit guard — no PHP shim
needed for those, since PHP's `new \Class()` doesn't chain.

Inner-class FQN bug fix: classFqn for non-JDK classes now strips `$`
matching Compiler::mangle. `BenchRecord$Point` was producing a
literal `\PHPJava\Aot\Generated\BenchRecord$Point` reference, parsing
as PHP variable interpolation; now `BenchRecord_Point` consistent
with the class declaration.

End-to-end record validation: bench/aot-record.php runs Java records
(BenchRecord$Point) through the full AOT pipeline:
  - Point.<init> as instance __construct
  - Point.equals/hashCode/toString as instance methods routing through
    the ObjectMethods bootstrap emit (commit 1aada77)
  - super(java/lang/Record).<init>() elided

All 4 record tests PASS:
  eqSame: 1 ✓  eqDiff: 0 ✓  hashCodeOk: 1 ✓  toStr: 'Point[x=3, y=5]' ✓

Real-library probe: 100% IR coverage on 4983 methods of commons-lang3
(up from 4197 — instance methods + __construct previously hidden by
the probe's static-only regex; fixed in the same commit).

Snapshot baseline updated — every fixture grew by ~135 bytes for
the now-emitted __construct body.