diff --git a/TODOS.md b/TODOS.md
index 445c5f6..eed6ab9 100644
--- a/TODOS.md
+++ b/TODOS.md
@@ -12,6 +12,26 @@
     parsers via a pool or per-thread storage, plugged in behind `open_parser`
     without touching call sites. A `TSParser`/`TSQueryCursor` is one-per-thread.
   - Full write-up: `GRAMMAR_CACHE_THREADING.md`.
+- [ ] Restore parallel grammar preload (BEHAVIOUR CHANGE in `edb2e64`).
+      `preload_grammars` used to `dlopen`+compile each grammar on its own thread;
+      it is now serial (`grammar()` per lang on the calling thread). The startup
+      dlopen cost is no longer overlapped.
+  - Why serial for now: `grammar()` holds `registry.mu` across the whole
+    dlopen+compile, so naively fanning out `grammar()` calls just queues every
+    worker on that one lock — zero real parallelism.
+  - To bring it back: move `build_grammar` (dlopen + `compile_query`, no shared
+    mutable state) OUTSIDE the lock; take `mu` only to (1) init the registry +
+    snapshot which langs are missing, and (2) briefly publish each finished
+    `^Grammar`. Restore the `thread`-based worker fan-out + `SPALL`
+    init/cleanup hooks that `edb2e64` deleted. Safe because glibc `dlopen`, the
+    heap allocator, and each thread's own temp allocator are all thread-safe.
+  - An `RW_Mutex` does NOT help the parallel-load side (writers stay exclusive);
+    it only helps a future multi-threaded *read* path (concurrent cache-hit
+    `grammar()` lookups during a parallel render). Consider it only if/when
+    render itself goes multi-threaded — then pair rlock hits with off-lock build
+    + a double-check-and-discard on the miss race (needs a `free_grammar`).
+  - Assumption to preserve: preload completes before any concurrent `grammar()`
+    use (render is single-threaded and runs after preload).
 - [ ] instead of `warnings: [dynamic]Error` we should use `warnings: [dynamic; 8]Error`
   - when limit reached, the template fails and stops rendering.
 - [ ] Brainstorm plugin/tool/(postccs/less/tailwind/sass) support.
diff --git a/markdown/highlight.odin b/markdown/highlight.odin
index 0ef5499..bebaa06 100644
--- a/markdown/highlight.odin
+++ b/markdown/highlight.odin
@@ -90,9 +90,12 @@ write_escaped :: proc(b: ^strings.Builder, s: string) {
 	}
 }
 
+// unescape_html returns a caller-owned string (free it with delete). When the
+// input has no entities it returns a clone; otherwise it returns the builder's
+// buffer. It must NOT destroy the builder — to_string aliases that buffer, so the
+// returned string owns it.
 unescape_html :: proc(s: string) -> string {
 	sb := strings.builder_make()
-	defer strings.builder_destroy(&sb)
 
 	start := 0
 	for i in 0 ..< len(s) {
@@ -119,7 +122,7 @@ unescape_html :: proc(s: string) -> string {
 		strings.write_string(&sb, replacement)
 		start = i + semi + 1
 	}
-	if start == 0 do return s
+	if start == 0 do return strings.clone(s)
 	if start < len(s) do strings.write_string(&sb, s[start:])
 	return strings.to_string(sb)
 }
@@ -136,6 +139,7 @@ highlight_block :: proc(code: string, lang: string, file_path: string) -> string
 	defer ts.parser_delete(parser)
 
 	raw_code := unescape_html(code)
+	defer delete(raw_code)
 	raw_c := strings.clone_to_cstring(raw_code)
 	defer delete(raw_c)
 
diff --git a/markdown/highlight_test.odin b/markdown/highlight_test.odin
index a37c7ad..87c0f34 100644
--- a/markdown/highlight_test.odin
+++ b/markdown/highlight_test.odin
@@ -81,3 +81,23 @@ test_highlight_code_wraps_css_block :: proc(t: ^testing.T) {
 		"block body should be highlighted",
 	)
 }
+
+// unescape_html must return a caller-owned allocation the caller frees exactly
+// once. The buggy version frees its builder internally (defer builder_destroy)
+// and returns the freed alias, so the caller's delete becomes a double-free that
+// the test runner's tracking allocator flags.
+@(test)
+test_unescape_html_entities_owned :: proc(t: ^testing.T) {
+	got := unescape_html("&lt;a&gt;")
+	defer delete(got)
+	testing.expect_value(t, got, "<a>")
+}
+
+// Same contract on the no-entity path: the buggy version returns the borrowed
+// input, so freeing it is a bad free (pointer not from the allocator).
+@(test)
+test_unescape_html_passthrough_owned :: proc(t: ^testing.T) {
+	got := unescape_html("plain text")
+	defer delete(got)
+	testing.expect_value(t, got, "plain text")
+}