x64 call stub was ABI-broken: fixed 0x20 frame left rsp misaligned at the
inner call (callee entry rsp ≡ 0, ABI requires ≡ 8) and, for 5+ args, wrote
stack args over the return address. Compute frame K ≡ 8 (mod 16), K ≥
0x20 + 8*stackArgs, so the callee sees a 16-aligned stack and stack args land
above the shadow window. Load register args as full 64-bit imm64 (was imm32,
which truncated pointers > 4 GiB). BuildCallStub now takes nuint[]; x86 range-
checks each arg against uint.MaxValue instead of silently truncating.
MarshalCache conflated managed and unmanaged width in one Size field: the
blittable path needs Unsafe.SizeOf<T> (bool = 1) while the marshal path needs
Marshal.SizeOf<T> (inline ByValTStr/ByValArray expand past the managed
pointer). Add MarshalSize; MemoryBase picks per TypeRequiresMarshal at all four
IO sites. Prevents PtrToStructure/StructureToPtr from over-reading/overwriting
the pinned scratch buffer (heap corruption on write).
Extract shared RPM/WPM into RpmHelper: honor partial reads (dead Array.Resize
removed), consistent write-return semantics; InProcessReader now guards
MainModule like ExternalReader.
Tests: x64 frame-alignment property + inline-marshal round-trip added (both
fail against the pre-fix code); existing x64 byte-expectation tests updated to
the new frame. Build clean, 100/100 pass.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
kbeandClaude Opus 4.8