diff --git a/benchmarks/encode_decode.exs b/benchmarks/encode_decode.exs
new file mode 100644
index 0000000..023a109
--- /dev/null
+++ b/benchmarks/encode_decode.exs
@@ -0,0 +1,186 @@
+# benchmarks/encode_decode.exs
+#
+# Isolated micro-benchmarks that decompose the Lua.encode!/2 + Lua.decode!/2
+# round-trip regression. On a nested map the native VM measured ~6x slower than
+# luerl 0.4.0 (18us -> 108us); every other scenario stayed within ~2x. This
+# script exists to attribute that 6x to a specific direction, container shape,
+# and size so the root cause is obvious rather than guessed.
+#
+# ---------------------------------------------------------------------------
+# How to read the output
+# ---------------------------------------------------------------------------
+# Each row reports total time for one operation and, crucially, the
+# per-element cost (per_elem_ns = total / element_count):
+#
+#   * per_elem_ns roughly CONSTANT as N grows  -> linear cost, nothing to see
+#   * per_elem_ns GROWS as N grows             -> super-linear (algorithmic)
+#     cost. That is the smoking gun -- e.g. O(n^2) list building, a decode that
+#     builds-then-reverses, or repeated rehashing/lookups per element.
+#
+# Comparing the `encode` rows to the `decode` rows attributes the round-trip
+# cost to a direction. Comparing shapes at equal N attributes it to a
+# container kind (list vs string-keyed map vs integer-keyed map vs records vs
+# depth). The nested_chain sweep isolates recursion/traversal depth from
+# fan-out.
+#
+# ---------------------------------------------------------------------------
+# How to run
+# ---------------------------------------------------------------------------
+# Inside any project that has `lua` as a dependency -- including the tv-labs/lua
+# repo itself, where `mix run` benches your working tree directly:
+#
+#     mix run benchmarks/encode_decode.exs
+#
+# To compare engines, run it once from a project pinned to lua 0.4.0 (luerl)
+# and once from one pinned to the native VM, then diff the per_elem_ns columns.
+# The script prints the loaded `lua` version in its header.
+
+defmodule EncDecBench do
+  # Sizes chosen to span ~3 orders of magnitude so a super-linear curve is
+  # unmistakable. Trim the large end if a shape gets too slow to be practical.
+  @sizes [8, 64, 512, 4096]
+  @depths [4, 16, 64, 256]
+
+  # Minimum wall-clock per timed batch; reps auto-scale to reach it so tiny
+  # ops aren't swamped by :timer.tc resolution. Median of @samples batches.
+  @min_batch_us 30_000
+  @samples 5
+
+  # ---- builders: each takes a size and returns a plain Elixir term ----
+  def int_list(n), do: Enum.to_list(1..n)
+  def float_list(n), do: Enum.map(1..n, &(&1 * 1.5))
+  def bool_list(n), do: Enum.map(1..n, fn i -> rem(i, 2) == 0 end)
+  def short_string_list(n), do: Enum.map(1..n, &"s#{&1}")
+
+  def long_string_list(n),
+    do: Enum.map(1..n, fn i -> String.duplicate("x", 200) <> Integer.to_string(i) end)
+
+  def string_map(n), do: Map.new(1..n, fn i -> {"key_#{i}", i} end)
+  def int_map(n), do: Map.new(1..n, fn i -> {i, i} end)
+
+  def record_list(n) do
+    Enum.map(1..n, fn i ->
+      %{"id" => i, "name" => "contact #{i}", "active" => rem(i, 2) == 0, "score" => i * 1.5}
+    end)
+  end
+
+  def nested_chain(0), do: %{"leaf" => true}
+  def nested_chain(d), do: %{"depth" => d, "child" => nested_chain(d - 1)}
+
+  # The exact composite from the PR benchmark, to anchor the 18us/108us number.
+  def original_nested do
+    %{
+      "name" => "benchmark contact",
+      "fields" => Map.new(1..20, fn i -> {"field_#{i}", "value #{i}"} end),
+      "tags" => Enum.map(1..50, &"tag-#{&1}"),
+      "meta" => %{"a" => 1, "b" => true, "c" => 3.14}
+    }
+  end
+
+  defp elem_count(term) when is_list(term), do: max(length(term), 1)
+  defp elem_count(term) when is_map(term), do: max(map_size(term), 1)
+  defp elem_count(_), do: 1
+
+  # ---- timing harness ----
+  defp time_batch(fun, reps) do
+    {us, _} = :timer.tc(fn -> Enum.each(1..reps, fn _ -> fun.() end) end)
+    us
+  end
+
+  defp calibrate(fun, reps) do
+    cond do
+      reps >= 5_000_000 -> reps
+      time_batch(fun, reps) >= @min_batch_us -> reps
+      true -> calibrate(fun, reps * 4)
+    end
+  end
+
+  # Microseconds per single call (median of @samples batches).
+  defp measure(fun) do
+    fun.()
+    reps = calibrate(fun, 1)
+    samples = for _ <- 1..@samples, do: time_batch(fun, reps) / reps
+    Enum.at(Enum.sort(samples), div(@samples, 2))
+  end
+
+  # ---- run ----
+  def run do
+    materialize? =
+      Code.ensure_loaded?(Lua.Table) and function_exported?(Lua.Table, :deep_cast, 1)
+
+    IO.puts("lua #{Application.spec(:lua, :vsn)} — encode!/decode! decomposition")
+    IO.puts("(decode+deep_cast column: #{if materialize?, do: "enabled", else: "N/A on this engine"})")
+    IO.puts(String.duplicate("=", 82))
+    header()
+
+    builders = [
+      {"int_list", &int_list/1},
+      {"float_list", &float_list/1},
+      {"bool_list", &bool_list/1},
+      {"short_string_list", &short_string_list/1},
+      {"long_string_list", &long_string_list/1},
+      {"string_map", &string_map/1},
+      {"int_map", &int_map/1},
+      {"record_list", &record_list/1}
+    ]
+
+    for {name, builder} <- builders do
+      for n <- @sizes, do: bench_shape(name, builder.(n), n, materialize?)
+      IO.puts(String.duplicate("-", 82))
+    end
+
+    IO.puts("nested chain (depth sweep) — isolates recursion/traversal from fan-out")
+    header()
+    for d <- @depths, do: bench_shape("nested_chain", nested_chain(d), d, materialize?)
+
+    IO.puts(String.duplicate("=", 82))
+    IO.puts("composite anchor — the PR's `original_nested` (matches the 18us/108us figure)")
+    header()
+    bench_shape("original_nested", original_nested(), 75, materialize?)
+  end
+
+  # Encodes from a FIXED base state (built once) so we time encoding, not
+  # Lua.new() (~80-107us, which would swamp small-N encode). The returned state
+  # is discarded, so the base never grows across reps.
+  defp bench_shape(name, term, n, materialize?) do
+    base = Lua.new()
+    count = elem_count(term)
+
+    enc_us = measure(fn -> Lua.encode!(base, term) end)
+    row("encode", name, n, enc_us, count)
+
+    {encoded, state} = Lua.encode!(base, term)
+    dec_us = measure(fn -> Lua.decode!(state, encoded) end)
+    row("decode", name, n, dec_us, count)
+
+    if materialize? do
+      mat_us = measure(fn -> Lua.Table.deep_cast(Lua.decode!(state, encoded)) end)
+      row("dec+cast", name, n, mat_us, count)
+    end
+  end
+
+  defp header do
+    IO.puts(
+      pad("op", 9) <>
+        pad("shape", 20) <>
+        lpad("N", 7) <>
+        lpad("total_us", 13) <>
+        lpad("per_elem_ns", 14)
+    )
+  end
+
+  defp row(op, shape, n, total_us, count) do
+    IO.puts(
+      pad(op, 9) <>
+        pad(shape, 20) <>
+        lpad(Integer.to_string(n), 7) <>
+        lpad(:erlang.float_to_binary(total_us, decimals: 2), 13) <>
+        lpad(:erlang.float_to_binary(total_us * 1000 / count, decimals: 1), 14)
+    )
+  end
+
+  defp pad(s, w), do: String.pad_trailing(s, w)
+  defp lpad(s, w), do: String.pad_leading(s, w)
+end
+
+EncDecBench.run()
diff --git a/lib/lua/vm/table.ex b/lib/lua/vm/table.ex
index 2bdb030..5116c5a 100644
--- a/lib/lua/vm/table.ex
+++ b/lib/lua/vm/table.ex
@@ -112,7 +112,12 @@ defmodule Lua.VM.Table do
   end
 
   defp split_from_map(table, data) do
-    Enum.reduce(data, table, fn {k, v}, acc -> put(acc, k, v) end)
+    # Sorted fold: integer keys come first, ascending (term order puts numbers
+    # before other keys), so contiguous appends hit put/3's O(1) array path
+    # instead of parking in the hash and draining via absorb_from_hash — an
+    # O(n^2) build for dense int maps whose iteration order is by hash (any map
+    # >32 entries). The sort is O(n log n), cheaper than the O(n^2) it replaces.
+    Enum.reduce(Enum.sort(data), table, fn {k, v}, acc -> put(acc, k, v) end)
   end
 
   @doc """