diff --git a/examples/ffi-02-small-struct.sx b/examples/ffi-02-small-struct.sx
new file mode 100644
index 0000000..5687551
--- /dev/null
+++ b/examples/ffi-02-small-struct.sx
@@ -0,0 +1,50 @@
+// Phase 0 baseline (PLAN-FFI.md step 0.2): small structs (≤16 bytes)
+// passed by value into a C `#foreign` fn and returned by value. Two
+// shapes that exercise different aggregate ABI paths today:
+//   Vec2  — 8 bytes,  two f32        (float register pair on AAPCS64)
+//   Vec4f — 16 bytes, four f32       (HFA — homogeneous float aggregate)
+//
+// 16-byte integer-only structs (e.g. `{ s64, s64 }`, `{ s32, s32, s32, s32 }`)
+// are *not* covered here: sx's `#foreign` decl currently lowers them
+// as `[2 x i64]` while the call site uses the struct type, tripping
+// the LLVM verifier. Repro pinned in `examples/issue-0036.sx`; once
+// that bug closes, fold those shapes back into this baseline.
+
+#import "modules/std.sx";
+
+// `#source` only — c_import would rewrite struct-typed params/returns
+// in the .h to *void (its "struct/opaque pointer → *void" default),
+// which would link but pass through the wrong ABI. The sx declarations
+// below match the C signatures exactly.
+#import c {
+    #source "vendors/ffi_structs/ffi_structs.c";
+};
+
+Vec2  :: struct { x: f32; y: f32; }
+Vec4f :: struct { x: f32; y: f32; z: f32; w: f32; }
+
+ffi_vec2_make :: (x: f32, y: f32) -> Vec2 #foreign;
+ffi_vec2_swap :: (v: Vec2)        -> Vec2 #foreign;
+ffi_vec2_sum  :: (v: Vec2)        -> f32  #foreign;
+
+ffi_vec4f_make    :: (x: f32, y: f32, z: f32, w: f32) -> Vec4f #foreign;
+ffi_vec4f_reverse :: (v: Vec4f)                       -> Vec4f #foreign;
+ffi_vec4f_sum     :: (v: Vec4f)                       -> f32   #foreign;
+
+main :: () -> s32 {
+    // ── Vec2 (8 bytes, float pair) ─────────────────────────────────
+    v := ffi_vec2_make(1.5, 2.5);
+    print("vec2 make = ({}, {})\n", v.x, v.y);
+    w := ffi_vec2_swap(v);
+    print("vec2 swap = ({}, {})\n", w.x, w.y);
+    print("vec2 sum  = {}\n", ffi_vec2_sum(v));
+
+    // ── Vec4f (16 bytes, HFA) ──────────────────────────────────────
+    f := ffi_vec4f_make(1.0, 2.0, 3.0, 4.0);
+    print("vec4f make = ({}, {}, {}, {})\n", f.x, f.y, f.z, f.w);
+    g := ffi_vec4f_reverse(f);
+    print("vec4f rev  = ({}, {}, {}, {})\n", g.x, g.y, g.z, g.w);
+    print("vec4f sum  = {}\n", ffi_vec4f_sum(f));
+
+    0;
+}
diff --git a/tests/expected/ffi-02-small-struct.exit b/tests/expected/ffi-02-small-struct.exit
new file mode 100644
index 0000000..573541a
--- /dev/null
+++ b/tests/expected/ffi-02-small-struct.exit
@@ -0,0 +1 @@
+0
diff --git a/tests/expected/ffi-02-small-struct.txt b/tests/expected/ffi-02-small-struct.txt
new file mode 100644
index 0000000..a29959e
--- /dev/null
+++ b/tests/expected/ffi-02-small-struct.txt
@@ -0,0 +1,6 @@
+vec2 make = (1.500000, 2.500000)
+vec2 swap = (2.500000, 1.500000)
+vec2 sum  = 4.000000
+vec4f make = (1.000000, 2.000000, 3.000000, 4.000000)
+vec4f rev  = (4.000000, 3.000000, 2.000000, 1.000000)
+vec4f sum  = 10.000000
diff --git a/vendors/ffi_structs/ffi_structs.c b/vendors/ffi_structs/ffi_structs.c
new file mode 100644
index 0000000..73728a2
--- /dev/null
+++ b/vendors/ffi_structs/ffi_structs.c
@@ -0,0 +1,29 @@
+#include "ffi_structs.h"
+
+Vec2 ffi_vec2_make(float x, float y) {
+    Vec2 r = { x, y };
+    return r;
+}
+
+Vec2 ffi_vec2_swap(Vec2 v) {
+    Vec2 r = { v.y, v.x };
+    return r;
+}
+
+float ffi_vec2_sum(Vec2 v) {
+    return v.x + v.y;
+}
+
+Vec4f ffi_vec4f_make(float x, float y, float z, float w) {
+    Vec4f r = { x, y, z, w };
+    return r;
+}
+
+Vec4f ffi_vec4f_reverse(Vec4f v) {
+    Vec4f r = { v.w, v.z, v.y, v.x };
+    return r;
+}
+
+float ffi_vec4f_sum(Vec4f v) {
+    return v.x + v.y + v.z + v.w;
+}
diff --git a/vendors/ffi_structs/ffi_structs.h b/vendors/ffi_structs/ffi_structs.h
new file mode 100644
index 0000000..35b7077
--- /dev/null
+++ b/vendors/ffi_structs/ffi_structs.h
@@ -0,0 +1,18 @@
+// FFI struct-marshalling baselines. Two shapes covered today:
+//   Vec2  — 8 bytes  (two f32)         — register pair, float path
+//   Vec4f — 16 bytes (four f32)        — homogeneous float aggregate (HFA)
+// Declared here so the .c has a header to include; the sx side
+// imports via `#source` only and re-declares the structs natively
+// (c_import currently rewrites struct-typed params/returns to *void,
+// which loses the by-value ABI).
+
+typedef struct { float x; float y; } Vec2;
+typedef struct { float x; float y; float z; float w; } Vec4f;
+
+Vec2  ffi_vec2_make  (float x, float y);
+Vec2  ffi_vec2_swap  (Vec2 v);
+float ffi_vec2_sum   (Vec2 v);
+
+Vec4f ffi_vec4f_make   (float x, float y, float z, float w);
+Vec4f ffi_vec4f_reverse(Vec4f v);
+float ffi_vec4f_sum    (Vec4f v);