Cache parsed components for 10x faster startup (2s → 200ms)

- Fix O(n²) postprocessing: compute_all_deps/io_refs/hash were called
  per-file (92x for sx app). Now deferred to single finalize_components()
  call after all files load.
- Add pickle cache in shared/sx/.cache/ keyed by file mtimes+sizes.
  Cache stores fully-processed Component/Island/Macro objects with deps,
  io_refs, and css_classes pre-computed. Closures stripped before pickle,
  rebuilt from global env after restore.
- Smart finalization: cached loads skip deps/io_refs recomputation
  (already in pickle), only recompute component hash.
- Fix ~sx-header → ~layouts/header ref in docs-content.sx

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

shared/sx/components.py

@@ -13,14 +13,18 @@ from .jinja_bridge import load_sx_dir, register_reload_callback, watch_sx_dir
 
 
 def load_shared_components() -> None:
-    """Register all shared s-expression components."""
+    """Register all shared s-expression components.
+
+    Defers finalization (deps/hash) so the calling app can load service
+    components before the single finalize pass.
+    """
     # Load SX libraries first — reader macros (#z3 etc.) must resolve
     # before any .sx file that uses them is parsed
     _load_sx_libraries()
     register_reload_callback(_load_sx_libraries)
 
     templates_dir = os.path.join(os.path.dirname(__file__), "templates")
-    load_sx_dir(templates_dir)
+    load_sx_dir(templates_dir, _finalize=False)
     watch_sx_dir(templates_dir)
 
 
@@ -32,4 +36,4 @@ def _load_sx_libraries() -> None:
         path = os.path.join(ref_dir, name)
         if os.path.exists(path):
             with open(path, encoding="utf-8") as f:
-                register_components(f.read())
+                register_components(f.read(), _defer_postprocess=True)