fix: strip nested mdata in grind preprocessing

This PR fixes an issue where `grind` would fail after multiple `norm_cast`
calls with the error "unexpected metadata found during internalization".

The `norm_cast` tactic adds mdata nodes to expressions, and when called
multiple times it creates nested mdata. The `eraseIrrelevantMData`
preprocessing function was using `.continue e` when stripping mdata,
which causes `Core.transform` to reconstruct the mdata node around the
visited children. By changing to `.visit e`, the inner expression is
passed back to `pre` for another round of processing, allowing all
nested mdata layers to be stripped.

Closes #11411

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>
diff --git a/src/Lean/Meta/Tactic/Grind/Util.lean b/src/Lean/Meta/Tactic/Grind/Util.lean
index 3422234..d93eba3 100644
--- a/src/Lean/Meta/Tactic/Grind/Util.lean
+++ b/src/Lean/Meta/Tactic/Grind/Util.lean
@@ -153,7 +153,7 @@
   let pre (e : Expr) := do
     match e with
     | .letE .. | .lam .. => return .done e
-    | .mdata _ e => return .continue e
+    | .mdata _ e => return .visit e
     | _ => return .continue e
   Core.transform e (pre := pre)
 
diff --git a/tests/lean/run/grind_nested_mdata.lean b/tests/lean/run/grind_nested_mdata.lean
new file mode 100644
index 0000000..f3f52a6
--- /dev/null
+++ b/tests/lean/run/grind_nested_mdata.lean
@@ -0,0 +1,59 @@
+/-!
+# Test that `grind` handles nested mdata from multiple `norm_cast` calls
+
+This is a regression test for https://github.com/leanprover/lean4/issues/11411
+
+The issue was that `norm_cast` adds mdata nodes to expressions, and when called
+multiple times it creates nested mdata. The `grind` preprocessing step
+`eraseIrrelevantMData` was only stripping one layer of mdata instead of all layers.
+-/
+
+-- Minimal WithTop definition
+def WithTop (α : Type u) := Option α
+
+namespace WithTop
+
+variable {α : Type u}
+
+@[coe] def some : α → WithTop α := Option.some
+
+instance : Coe α (WithTop α) := ⟨some⟩
+
+instance : NatCast (WithTop Nat) where
+  natCast n := some n
+
+end WithTop
+
+-- AtLeastTwo infrastructure (from Mathlib.Data.Nat.Init)
+class Nat.AtLeastTwo (n : Nat) : Prop where
+  prop : n ≥ 2
+
+instance Nat.instAtLeastTwo : Nat.AtLeastTwo (n + 2) where
+  prop := Nat.le_add_left 2 n
+
+-- OfNat instance for numeric literals ≥ 2 (from Mathlib.Data.Nat.Cast.Defs)
+instance instOfNatAtLeastTwo {R : Type u} [NatCast R] (n : Nat) [Nat.AtLeastTwo n] : OfNat R n where
+  ofNat := n
+
+-- The key norm_cast lemma (from Mathlib.Data.Nat.Cast.Defs)
+@[simp, norm_cast]
+theorem Nat.cast_ofNat {R : Type u} {n : Nat} [NatCast R] [Nat.AtLeastTwo n] :
+    (Nat.cast (no_index (OfNat.ofNat n)) : R) = no_index (OfNat.ofNat n) := rfl
+
+variable {Foo : WithTop Nat → Prop} {Bar : Prop}
+
+-- Works: grind without norm_cast
+example (foo_imp_bar : ∀ x, Foo x → Bar) (h : Foo 2) : Bar := by
+  grind
+
+-- Works: grind after one norm_cast
+example (foo_imp_bar : ∀ x, Foo x → Bar) (h : Foo 2) : Bar := by
+  norm_cast at h
+  grind
+
+-- Previously failed: grind after two norm_casts
+-- This used to fail with "unexpected metadata found during internalization"
+example (foo_imp_bar : ∀ x, Foo x → Bar) (h : Foo 2) : Bar := by
+  norm_cast at h
+  norm_cast at h
+  grind