docs(binary-protocol): document the unboxed NodeRef content field

advanced/binary-protocol.mdx

@@ -654,7 +654,7 @@ pub enum NodeStr<'a> {
 pub struct NodeRef<'a> {
     pub tag: NodeStr<'a>,          // Borrowed or inline
     pub attrs: AttrsRef<'a>,       // Vec<(NodeStr<'a>, ValueRef<'a>)>
-    pub content: Option<Box<NodeContentRef<'a>>>,
+    pub content: Option<NodeContentRef<'a>>,
 }
 
 pub enum NodeContentRef<'a> {
@@ -670,6 +670,14 @@ pub enum NodeContentRef<'a> {
   `NodeStr` replaces the previous `Cow<'a, str>` used in `NodeRef`, `AttrsRef`, `ValueRef`, and `NodeContentRef`. The key difference is that the `Owned` variant uses `CompactString` (inline up to 24 bytes) instead of `String` (always heap-allocated), reducing allocation pressure for the many short protocol strings that can't be statically interned.
 </Note>
 
+<Note>
+  `NodeRef::content` dropped its `Box` — it is now `Option<NodeContentRef<'a>>` rather than `Option<Box<NodeContentRef<'a>>>`. Code matching on it via `.as_deref()` should switch to `.as_ref()`; a `Box<NodeContentRef<'_>>` you're holding directly derefs the same as before.
+
+  The box existed only to keep `NodeRef` pointer-sized, but for the commonest node in a fanout — a leaf with `Cow::Borrowed` byte content, which already allocates nothing — the box was the *only* allocation. Removing it cuts allocations per decode by 40% on a fanout (45 → 27) and drops native decode time by 11.2% on both the fanout and large-stanza benchmarks. `size_of::<NodeRef>()` grows from 48 to 72 bytes in exchange, which does not regress any measured shape, including one with no content to begin with.
+
+  `NodeContentRef::Nodes` keeps its own `Box<NodeVec<'a>>` — that recursion is what makes the type coherent, and removing it isn't part of this change.
+</Note>
+
 Location: `wacore/binary/src/node.rs:10-106`, `465-469`, `437-441`
 
 ### Node nesting depth cap
@@ -936,6 +944,12 @@ The `spilled()` method on `SmallVec` reports whether a given node's attrs overfl
 
 Location: `wacore/binary/src/node.rs`, `wacore/binary/tests/attrs_inline_alloc.rs`
 
+### Unboxed NodeRef content
+
+`NodeRef::content` (`Option<NodeContentRef<'a>>`) has no `Box` around the inner value — see the note under [Zero-copy decoding](#zero-copy-decoding). On `bench_unmarshal_fanout`, this drops allocations per decode from 45 to 27 and instructions by 9.5% under callgrind (the instrument CodSpeed uses), alongside an 11.2% native wall-time improvement on both the fanout and large-stanza benchmarks. `bench_unmarshal_small`, which carries no content, is unaffected either way — the gain tracks how many nodes in the tree carry content, not a fixed per-node cost.
+
+Location: `wacore/binary/src/node.rs`, `wacore/binary/benches/binary_benchmark.rs`
+
 ## Common protocol patterns
 
 ### IQ (info/query) stanzas