diff --git a/crates/aura-engine/src/blueprint.rs b/crates/aura-engine/src/blueprint.rs
index cfe400d..bc5f68f 100644
--- a/crates/aura-engine/src/blueprint.rs
+++ b/crates/aura-engine/src/blueprint.rs
@@ -872,6 +872,69 @@ mod tests {
         assert_eq!(space[2].kind, ScalarKind::F64);
     }
 
+    /// E2E (issue #34): the C23/#31 mirror invariant *under composite nesting*.
+    /// `param_space()` (via `collect_params`) duplicates `lower_items`' depth-
+    /// first traversal rather than sharing it, so the two orders must stay in
+    /// lockstep. The single-level mirror test
+    /// (`param_space_mirrors_compiled_flat_node_param_order`) never compiles a
+    /// composite whose interior holds *another* composite; the nested
+    /// `param_space` order test never compiles. This closes that gap: it
+    /// compiles a `strategy → { fast_slow → [Sma, Sma, Sub], LinComb }` nest and
+    /// asserts the aggregated space lines up, kind-by-slot, with the compiled
+    /// flat-node param order — so a future inliner reorder that desynced the two
+    /// projections *only under nesting* would fail here, not slip through.
+    #[test]
+    fn param_space_mirrors_compiled_flat_node_param_order_under_nesting() {
+        use aura_std::{LinComb, Sma, Sub};
+        // inner composite "fast_slow": two SMAs + a Sub (same nest as the
+        // isolated path-qualification test above)
+        let fast_slow = Composite::new(
+            "fast_slow",
+            vec![Sma::new(2).into(), Sma::new(4).into(), Sub::new().into()],
+            vec![
+                Edge { from: 0, to: 2, slot: 0, from_field: 0 },
+                Edge { from: 1, to: 2, slot: 1, from_field: 0 },
+            ],
+            vec![vec![Target { node: 0, slot: 0 }, Target { node: 1, slot: 0 }]],
+            OutPort { node: 2, field: 0 },
+        );
+        // outer composite "strategy": the inner composite + a LinComb([1,-1])
+        let strategy = Composite::new(
+            "strategy",
+            vec![BlueprintNode::Composite(fast_slow), LinComb::new(vec![1.0, -1.0]).into()],
+            vec![],
+            vec![vec![Target { node: 0, slot: 0 }]],
+            OutPort { node: 0, field: 0 },
+        );
+        let bp = Blueprint::new(vec![BlueprintNode::Composite(strategy)], vec![], vec![]);
+
+        // the aggregated, path-qualified projection (borrows; take it first since
+        // compile() consumes self — same ordering as the single-level mirror test)
+        let space = bp.param_space();
+
+        // the same blueprint, compiled to its flat node array; each flat node's
+        // own declared params, concatenated in flat-node order
+        let (flat_nodes, _sources, _edges) = bp.compile().expect("nested composite compiles");
+        let from_compilat: Vec<ParamSpec> =
+            flat_nodes.iter().flat_map(|n| n.schema().params).collect();
+
+        // same count, same per-slot kind, same order — the nested projection
+        // mirrors the compilation (names differ: param_space path-qualifies, the
+        // raw node does not, so compare on the load-bearing axis, kind-by-slot)
+        assert_eq!(space.len(), from_compilat.len(), "param count must match the compilat");
+        assert_eq!(
+            space.iter().map(|p| p.kind).collect::<Vec<_>>(),
+            from_compilat.iter().map(|p| p.kind).collect::<Vec<_>>(),
+            "per-slot param kinds must line up with the compiled flat-node order, under nesting",
+        );
+        // pin the concrete shape: two Sma lengths (I64), Sub none, two LinComb
+        // weights (F64)
+        assert_eq!(
+            space.iter().map(|p| p.kind).collect::<Vec<_>>(),
+            [ScalarKind::I64, ScalarKind::I64, ScalarKind::F64, ScalarKind::F64],
+        );
+    }
+
     #[test]
     fn top_level_leaf_params_are_unqualified() {
         use aura_std::Sma;