trie/bintrie: use bitarray for path encoding + fix serialization issues (#34772)

Replace 1-byte-per-bit path encoding with bit-packed `BitArray`, reducing DB key size by 8x Benchmark (sparse single-leaf write, M3 Pro): ``` │ Before (1B/bit) │ After (BitArray) │ │ sec/op │ sec/op vs base │ CollectNodesSparseWrite-11 10.50µ ± 1% 9.78µ ± 1% -6.86% │ B/op │ B/op vs base │ CollectNodesSparseWrite-11 5.50Ki ± 0% 5.09Ki ± 0% -7.38% │ allocs/op │ allocs vs base │ CollectNodesSparseWrite-11 67 ± 0% 58 ± 0% -13.43% ``` --------- Co-authored-by: Guillaume Ballet <3272758+gballet@users.noreply.github.com>
2026-06-25 16:06:18 +00:00 · 2026-06-23 03:26:17 +07:00 · 2026-06-23 03:26:17 +07:00 · 73eeee65bf
commit 73eeee65bf
parent 023a573732
12 changed files with 1080 additions and 160 deletions
--- a/trie/bintrie/binary_node_test.go
+++ b/trie/bintrie/binary_node_test.go
@ -40,7 +40,10 @@ func TestSerializeDeserializeInternalNode(t *testing.T) {
 	s.root = rootRef

 	// Serialize the node — grouped format at groupDepth=1:
-	// [type(1)][groupDepth(1)][bitmap(1)][leftHash(32)][rightHash(32)] = 67 bytes
+	// [type(1)][groupDepth(1)][bitmap(1)][depths(1)][leftHash(32)][rightHash(32)] = 68 bytes.
+	// Both children are at depthOffset=groupDepth=1 (the bottom of the 1-level
+	// group). Each depth is stored as (offset-1)=0 in 3 bits, so the two entries
+	// pack into a single byte 0x00.
 	serialized := s.serializeNode(rootRef, 1)

 	if serialized[0] != nodeTypeInternal {
@ -50,7 +53,7 @@ func TestSerializeDeserializeInternalNode(t *testing.T) {
 		t.Errorf("Expected groupDepth byte to be 1, got %d", serialized[1])
 	}

-	expectedLen := NodeTypeBytes + 1 + 1 + 2*HashSize // type + groupDepth + bitmap + 2 hashes = 67
+	expectedLen := NodeTypeBytes + 1 + 1 + 1 + 2*HashSize // type + groupDepth + bitmap + packed depths + 2 hashes = 68
 	if len(serialized) != expectedLen {
 		t.Errorf("Expected serialized length to be %d, got %d", expectedLen, len(serialized))
 	}
@ -60,7 +63,13 @@ func TestSerializeDeserializeInternalNode(t *testing.T) {
 		t.Errorf("Expected bitmap byte 0xc0, got 0x%02x", serialized[2])
 	}

-	hashesStart := NodeTypeBytes + 1 + 1
+	depthsStart := NodeTypeBytes + 1 + 1
+	// Two depth offsets of 1 → stored as (1-1)=0 each → packed byte 0x00.
+	if serialized[depthsStart] != 0x00 {
+		t.Errorf("Expected packed depth byte 0x00, got 0x%02x", serialized[depthsStart])
+	}
+
+	hashesStart := depthsStart + 1
 	if !bytes.Equal(serialized[hashesStart:hashesStart+HashSize], leftHash[:]) {
 		t.Error("Left hash not found at expected position")
 	}
@ -244,29 +253,29 @@ func TestKeyToPath(t *testing.T) {
 		{
 			name:     "depth 0",
 			depth:    0,
-			key:      []byte{0x80}, // 10000000 in binary
-			expected: []byte{1},
+			key:      []byte{0x80},    // 10000000 in binary
+			expected: []byte{0x80, 1}, // 1 bit "1", left-aligned, + length byte 1
 			wantErr:  false,
 		},
 		{
 			name:     "depth 7",
 			depth:    7,
-			key:      []byte{0xFF}, // 11111111 in binary
-			expected: []byte{1, 1, 1, 1, 1, 1, 1, 1},
+			key:      []byte{0xFF},    // 11111111 in binary
+			expected: []byte{0xFF, 8}, // 8-bit value 0xFF + length byte 8
 			wantErr:  false,
 		},
 		{
 			name:     "depth crossing byte boundary",
 			depth:    10,
-			key:      []byte{0xFF, 0x00}, // 11111111 00000000 in binary
-			expected: []byte{1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0},
+			key:      []byte{0xFF, 0x00},     // 11111111 00000000 in binary
+			expected: []byte{0xFF, 0x00, 11}, // top 11 bits "11111111 000", left-aligned, + length byte 11
 			wantErr:  false,
 		},
 		{
 			name:     "max valid depth",
 			depth:    StemSize*8 - 1,
 			key:      make([]byte, HashSize),
-			expected: make([]byte, StemSize*8),
+			expected: append(make([]byte, StemSize), StemSize*8), // 248 bits of zeros + length byte 248
 			wantErr:  false,
 		},
 		{
--- a/trie/bintrie/bitarray.go
+++ b/trie/bintrie/bitarray.go
@ -0,0 +1,160 @@
+// Copyright 2026 The go-ethereum Authors
+// This file is part of go-ethereum.
+//
+// go-ethereum is free software: you can redistribute it and/or modify
+// it under the terms of the GNU General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// go-ethereum is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU General Public License for more details.
+//
+// You should have received a copy of the GNU General Public License
+// along with go-ethereum. If not, see <http://www.gnu.org/licenses/>.
+package bintrie
+
+// BitArray represents a trie path: the most significant `len` bits of a key,
+// packed big-endian and MSB-first. Bit i (0 = most significant) lives at
+// bytes[i/8] in mask 1<<(7-i%8). All bits at positions >= len are kept zero so
+// that two paths are byte-equal iff they are logically equal.
+//
+// This mirrors the on-disk key layout, so path manipulation is plain slicing
+// and copying: no shifting or endianness conversion is required. The maximum
+// length is 248 bits (a 31-byte trie stem), and is a uint8 so the spare bits in
+// the final byte are always available.
+type BitArray struct {
+	len   uint8
+	bytes [32]byte
+}
+
+// NewBitArray creates a bit array of the given length whose bits are the `length`
+// least-significant bits of val, read most-significant-first. Used by tests to
+// build expected paths; the value is interpreted as a number, not raw bytes.
+func NewBitArray(length uint8, val uint64) BitArray {
+	var b BitArray
+	b.len = length
+	for p := uint8(0); p < length; p++ {
+		if (val>>(length-1-p))&1 == 1 {
+			b.bytes[p/8] |= 1 << (7 - p%8)
+		}
+	}
+	return b
+}
+
+// Len returns the number of used bits.
+func (b *BitArray) Len() uint8 {
+	return b.len
+}
+
+// Bytes returns the packed big-endian, MSB-first representation. Bits beyond
+// len are zero.
+func (b *BitArray) Bytes() [32]byte {
+	return b.bytes
+}
+
+// AppendBit sets the bit array to x with a single bit appended, and returns the
+// receiver. Safe when b and x alias the same value.
+func (b *BitArray) AppendBit(x *BitArray, bit uint8) *BitArray {
+	*b = *x
+	if bit&1 == 1 {
+		// Position b.len is guaranteed zero by the all-bits-beyond-len-are-zero
+		// invariant, so a 1 only needs setting; a 0 is already in place.
+		b.bytes[b.len/8] |= 1 << (7 - b.len%8)
+	}
+	b.len++
+	return b
+}
+
+// MSBs sets the bit array to the most significant n bits of x and returns the
+// receiver. If n >= x.len it is an exact copy of x. Think of it as x[:n].
+func (b *BitArray) MSBs(x *BitArray, n uint8) *BitArray {
+	*b = *x
+	if n < b.len {
+		b.len = n
+		b.maskTail()
+	}
+	return b
+}
+
+// Equal reports whether two bit arrays hold the same path.
+func (b *BitArray) Equal(x *BitArray) bool {
+	return b.len == x.len && b.bytes == x.bytes
+}
+
+// SetBytes sets the bit array to the most significant `length` bits of data,
+// interpreted as big-endian bytes, and returns the receiver. At most 32 bytes
+// of data are read; bits beyond length are zeroed.
+func (b *BitArray) SetBytes(length uint8, data []byte) *BitArray {
+	b.bytes = [32]byte{}
+	copy(b.bytes[:], data)
+	b.len = length
+	b.maskTail()
+	return b
+}
+
+// SetBit sets the bit array to a single bit and returns the receiver.
+func (b *BitArray) SetBit(bit uint8) *BitArray {
+	b.bytes = [32]byte{}
+	b.len = 1
+	if bit&1 == 1 {
+		b.bytes[0] = 0x80
+	}
+	return b
+}
+
+// Copy returns a value copy of the bit array.
+func (b *BitArray) Copy() BitArray {
+	return *b
+}
+
+// Set sets the bit array to the same value as x and returns the receiver.
+func (b *BitArray) Set(x *BitArray) *BitArray {
+	*b = *x
+	return b
+}
+
+// KeyBytes returns the path-to-DB-key encoding: the active bytes (the
+// left-aligned MSB-first prefix) followed by a single trailing byte holding the
+// bit-length. The trailing length disambiguates paths whose active bytes
+// coincide (e.g. 1-bit "1" packs to [0x80, 0x01] and 8-bit "10000000" to
+// [0x80, 0x08]). The empty path encodes as no bytes.
+func (b *BitArray) KeyBytes() []byte {
+	if b.len == 0 {
+		return nil
+	}
+	bc := (int(b.len) + 7) / 8
+	res := make([]byte, bc+1)
+	copy(res[:bc], b.bytes[:bc])
+	res[bc] = b.len
+	return res
+}
+
+// PutKeyBytes writes the key encoding (active bytes followed by length byte)
+// into dst and returns the populated sub-slice. The empty path returns dst[:0]
+// without touching dst. For non-empty paths dst must have len >= 33 (32 packed
+// bytes for 248 bits + 1 length byte).
+func (b *BitArray) PutKeyBytes(dst []byte) []byte {
+	if b.len == 0 {
+		return dst[:0]
+	}
+	bc := (int(b.len) + 7) / 8
+	_ = dst[bc] // bounds check hint
+	copy(dst[:bc], b.bytes[:bc])
+	dst[bc] = b.len
+	return dst[:bc+1]
+}
+
+// maskTail zeroes every bit at a position >= len, preserving the invariant that
+// equal paths are byte-equal.
+func (b *BitArray) maskTail() {
+	full := int(b.len / 8)
+	if rem := b.len % 8; rem != 0 {
+		b.bytes[full] &= byte(0xFF) << (8 - rem)
+		full++
+	}
+	for i := full; i < len(b.bytes); i++ {
+		b.bytes[i] = 0
+	}
+}
--- a/trie/bintrie/bitarray_test.go
+++ b/trie/bintrie/bitarray_test.go
@ -0,0 +1,205 @@
+package bintrie
+
+import (
+	"bytes"
+	"testing"
+)
+
+// ba builds a BitArray with the given length and leading bytes, for use as an
+// expected value. Remaining bytes are zero.
+func ba(length uint8, lead ...byte) BitArray {
+	var b BitArray
+	b.len = length
+	copy(b.bytes[:], lead)
+	return b
+}
+
+func TestNewBitArray(t *testing.T) {
+	tests := []struct {
+		name   string
+		length uint8
+		val    uint64
+		want   BitArray
+	}{
+		{"empty", 0, 0, ba(0)},
+		{"single 1", 1, 1, ba(1, 0x80)},
+		{"single 0", 1, 0, ba(1, 0x00)},
+		{"101", 3, 0b101, ba(3, 0xA0)},
+		{"full byte", 8, 0xFF, ba(8, 0xFF)},
+		{"ten bits", 10, 0x3FF, ba(10, 0xFF, 0xC0)},
+		{"high bits ignored beyond length", 3, 0b11101, ba(3, 0xA0)},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := NewBitArray(tt.length, tt.val)
+			if !got.Equal(&tt.want) {
+				t.Errorf("NewBitArray(%d, %#x) = %x (len %d), want %x (len %d)",
+					tt.length, tt.val, got.bytes, got.len, tt.want.bytes, tt.want.len)
+			}
+		})
+	}
+}
+
+func TestSetBytes(t *testing.T) {
+	tests := []struct {
+		name   string
+		length uint8
+		data   []byte
+		want   BitArray
+	}{
+		{"empty", 0, []byte{0xFF}, ba(0)},
+		{"full byte", 8, []byte{0xAB}, ba(8, 0xAB)},
+		{"top 4 bits", 4, []byte{0xFF}, ba(4, 0xF0)},
+		{"11 bits masks tail", 11, []byte{0xFF, 0xFF}, ba(11, 0xFF, 0xE0)},
+		{"data longer than length", 4, []byte{0xFF, 0xFF}, ba(4, 0xF0)},
+		{"data shorter than length", 16, []byte{0xAB}, ba(16, 0xAB, 0x00)},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := new(BitArray).SetBytes(tt.length, tt.data)
+			if !got.Equal(&tt.want) {
+				t.Errorf("SetBytes(%d, %x) = %x (len %d), want %x (len %d)",
+					tt.length, tt.data, got.bytes, got.len, tt.want.bytes, tt.want.len)
+			}
+		})
+	}
+}
+
+func TestSetBytesFull(t *testing.T) {
+	data := bytes.Repeat([]byte{0xFF}, 32)
+	got := new(BitArray).SetBytes(248, data)
+	want := ba(248)
+	for i := 0; i < 31; i++ {
+		want.bytes[i] = 0xFF
+	}
+	if !got.Equal(&want) {
+		t.Errorf("SetBytes(248, 0xFF*32): byte 31 must be zeroed; got %x", got.bytes)
+	}
+}
+
+func TestMSBs(t *testing.T) {
+	x := new(BitArray).SetBytes(16, []byte{0xAB, 0xCD})
+	tests := []struct {
+		name string
+		n    uint8
+		want BitArray
+	}{
+		{"prefix byte", 8, ba(8, 0xAB)},
+		{"prefix nibble", 4, ba(4, 0xA0)},
+		{"zero", 0, ba(0)},
+		{"n equals len", 16, ba(16, 0xAB, 0xCD)},
+		{"n exceeds len copies x", 20, ba(16, 0xAB, 0xCD)},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got := new(BitArray).MSBs(x, tt.n)
+			if !got.Equal(&tt.want) {
+				t.Errorf("MSBs(x, %d) = %x (len %d), want %x (len %d)",
+					tt.n, got.bytes, got.len, tt.want.bytes, tt.want.len)
+			}
+		})
+	}
+}
+
+func TestAppendBit(t *testing.T) {
+	// Build "101" one bit at a time from empty.
+	var p BitArray
+	for _, bit := range []uint8{1, 0, 1} {
+		p.AppendBit(&p, bit) // receiver aliases argument
+	}
+	if want := ba(3, 0xA0); !p.Equal(&want) {
+		t.Fatalf("append 1,0,1 = %x (len %d), want %x (len 3)", p.bytes, p.len, want.bytes)
+	}
+
+	// Append across a byte boundary: 8 ones then a 1 → 9 bits.
+	var q BitArray
+	for i := 0; i < 9; i++ {
+		q.AppendBit(&q, 1)
+	}
+	if want := ba(9, 0xFF, 0x80); !q.Equal(&want) {
+		t.Fatalf("append nine 1s = %x (len %d), want %x (len 9)", q.bytes, q.len, want.bytes)
+	}
+
+	// Appending to a copy must not mutate the source.
+	src := new(BitArray).SetBytes(4, []byte{0xF0})
+	child := *src
+	child.AppendBit(&child, 0)
+	if want := ba(4, 0xF0); !src.Equal(&want) {
+		t.Errorf("source mutated by append on copy: %x", src.bytes)
+	}
+	if want := ba(5, 0xF0); !child.Equal(&want) {
+		t.Errorf("append 0 = %x (len %d), want %x (len 5)", child.bytes, child.len, want.bytes)
+	}
+}
+
+func TestSetBit(t *testing.T) {
+	if got, want := new(BitArray).SetBit(1), ba(1, 0x80); !got.Equal(&want) {
+		t.Errorf("SetBit(1) = %x (len %d), want %x", got.bytes, got.len, want.bytes)
+	}
+	if got, want := new(BitArray).SetBit(0), ba(1, 0x00); !got.Equal(&want) {
+		t.Errorf("SetBit(0) = %x (len %d), want %x", got.bytes, got.len, want.bytes)
+	}
+}
+
+func TestEqual(t *testing.T) {
+	a := NewBitArray(3, 0b101)
+	b := NewBitArray(3, 0b101)
+	if !a.Equal(&b) {
+		t.Error("equal arrays reported unequal")
+	}
+	// Same active bytes, different length must be unequal.
+	c := NewBitArray(2, 0b10) // "10" -> byte 0x80, len 2
+	d := ba(3, c.bytes[0])    // same byte, len 3
+	if c.Equal(&d) {
+		t.Error("arrays with different length reported equal")
+	}
+}
+
+func TestKeyBytesRoundTrip(t *testing.T) {
+	tests := []struct {
+		name   string
+		length uint8
+		data   []byte
+		want   []byte // expected KeyBytes output
+	}{
+		{"empty", 0, nil, nil},
+		{"one bit", 1, []byte{0x80}, []byte{0x80, 1}},
+		{"full byte", 8, []byte{0x80}, []byte{0x80, 8}},
+		{"eleven bits", 11, []byte{0xFF, 0xFF}, []byte{0xFF, 0xE0, 11}},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			src := new(BitArray).SetBytes(tt.length, tt.data)
+			key := src.KeyBytes()
+			if !bytes.Equal(key, tt.want) {
+				t.Fatalf("KeyBytes() = %x, want %x", key, tt.want)
+			}
+
+			// PutKeyBytes must agree with KeyBytes.
+			var buf [33]byte
+			if put := src.PutKeyBytes(buf[:]); !bytes.Equal(put, tt.want) {
+				t.Fatalf("PutKeyBytes() = %x, want %x", put, tt.want)
+			}
+
+			// Re-parse the active bytes and confirm the path round-trips.
+			if tt.length == 0 {
+				return
+			}
+			lengthByte := key[len(key)-1]
+			reparsed := new(BitArray).SetBytes(lengthByte, key[:len(key)-1])
+			if !reparsed.Equal(src) {
+				t.Fatalf("round-trip mismatch: %x (len %d) != %x (len %d)",
+					reparsed.bytes, reparsed.len, src.bytes, src.len)
+			}
+		})
+	}
+}
+
+func TestCopyIsIndependent(t *testing.T) {
+	src := new(BitArray).SetBytes(8, []byte{0xAB})
+	cp := src.Copy()
+	cp.AppendBit(&cp, 1)
+	if want := ba(8, 0xAB); !src.Equal(&want) {
+		t.Errorf("Copy not independent: source became %x (len %d)", src.bytes, src.len)
+	}
+}
--- a/trie/bintrie/format_test.go
+++ b/trie/bintrie/format_test.go
@ -0,0 +1,445 @@
+// Copyright 2026 go-ethereum Authors
+// This file is part of the go-ethereum library.
+//
+// The go-ethereum library is free software: you can redistribute it and/or modify
+// it under the terms of the GNU Lesser General Public License as published by
+// the Free Software Foundation, either version 3 of the License, or
+// (at your option) any later version.
+//
+// The go-ethereum library is distributed in the hope that it will be useful,
+// but WITHOUT ANY WARRANTY; without even the implied warranty of
+// MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+// GNU Lesser General Public License for more details.
+//
+// You should have received a copy of the GNU Lesser General Public License
+// along with the go-ethereum library. If not, see <http://www.gnu.org/licenses/>.
+
+package bintrie
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"testing"
+
+	"github.com/ethereum/go-ethereum/common"
+	"github.com/ethereum/go-ethereum/trie"
+	"github.com/ethereum/go-ethereum/trie/trienode"
+)
+
+// TestRootHashMatchesReadBackHash pins the round-trip invariant: the root
+// hash a Commit advertises must be exactly the value a fresh reader computes
+// from the on-disk blob. Before Option B the writer produced a natural-depth
+// hash while DeserializeAndHash produced an extended-depth hash, so the two
+// disagreed for any non-trivial subtree — this test failed. With the
+// per-entry depth byte, the reader rebuilds the natural-shape tree and the
+// hashes match for every groupDepth and every divergence bit.
+func TestRootHashMatchesReadBackHash(t *testing.T) {
+	for groupDepth := 1; groupDepth <= MaxGroupDepth; groupDepth++ {
+		// divergeBit ∈ [0, groupDepth-1] places the two stems at natural
+		// depth (divergeBit+1) within the root group; we want to exercise
+		// every depth-offset value the new format must handle.
+		for divergeBit := 0; divergeBit < groupDepth; divergeBit++ {
+			t.Run(fmt.Sprintf("gd=%d/diverge=%d", groupDepth, divergeBit), func(t *testing.T) {
+				tr := &BinaryTrie{
+					store:      newNodeStore(),
+					tracer:     trie.NewPrevalueTracer(),
+					groupDepth: groupDepth,
+				}
+				stemL, stemR := stemsDivergingAt(divergeBit)
+				if err := tr.store.Insert(stemL, oneKey[:], nil); err != nil {
+					t.Fatalf("Insert stemL: %v", err)
+				}
+				if err := tr.store.Insert(stemR, twoKey[:], nil); err != nil {
+					t.Fatalf("Insert stemR: %v", err)
+				}
+
+				natural := tr.Hash()
+				_, ns := tr.Commit(false)
+				rootNode, ok := ns.Nodes[""]
+				if !ok {
+					t.Fatalf("Commit produced no root blob (path \"\")")
+				}
+
+				readBack, err := DeserializeAndHash(rootNode.Blob, 0)
+				if err != nil {
+					t.Fatalf("DeserializeAndHash: %v", err)
+				}
+				if natural != readBack {
+					t.Fatalf("round-trip hash mismatch:\n"+
+						"  tr.Hash()                    = %x\n"+
+						"  DeserializeAndHash(rootBlob) = %x\n"+
+						"the parent's stored root hash cannot be reproduced from its own blob",
+						natural, readBack)
+				}
+			})
+		}
+	}
+}
+
+// TestMultiStemMixedDepths inserts four stems that diverge at different
+// depths within a single groupDepth=5 group, then round-trips the trie
+// through Commit + fresh-read. Verifies that every stem is retrievable by
+// key after reload — exercises the new format with several depth-offset
+// values in the same blob (1, 2, 3, 4) and confirms attachInGroup builds
+// the natural-shape tree correctly.
+func TestMultiStemMixedDepths(t *testing.T) {
+	const groupDepth = 5
+
+	// Each stem diverges from `0x00…00` at a different bit, so naturally:
+	//  - stem at bit-0 divergence → depth 1
+	//  - stem at bit-1 divergence → depth 2
+	//  - stem at bit-2 divergence → depth 3
+	//  - stem at bit-3 divergence → depth 4
+	stems := [][]byte{
+		zeroKey[:],
+		bitFlipStem(0), // diverge at bit 0
+		bitFlipStem(1), // diverge at bit 1 (prefix "0" matches stem 0)
+		bitFlipStem(2), // diverge at bit 2 (prefix "00")
+		bitFlipStem(3), // diverge at bit 3 (prefix "000")
+	}
+	values := []common.Hash{oneKey, twoKey, threeKey, fourKey, ffKey}
+
+	tr := &BinaryTrie{
+		store:      newNodeStore(),
+		tracer:     trie.NewPrevalueTracer(),
+		groupDepth: groupDepth,
+	}
+	for i, stem := range stems {
+		if err := tr.store.Insert(stem, values[i][:], nil); err != nil {
+			t.Fatalf("Insert stem %d: %v", i, err)
+		}
+	}
+
+	before := tr.Hash()
+	_, ns := tr.Commit(false)
+	rootBlob, ok := ns.Nodes[""]
+	if !ok {
+		t.Fatalf("no root blob in NodeSet")
+	}
+	readBack, err := DeserializeAndHash(rootBlob.Blob, 0)
+	if err != nil {
+		t.Fatalf("DeserializeAndHash: %v", err)
+	}
+	if before != readBack {
+		t.Fatalf("hash mismatch: tr.Hash()=%x DeserializeAndHash(rootBlob)=%x", before, readBack)
+	}
+
+	// Reload the root blob into a fresh store and confirm structure.
+	fresh := newNodeStore()
+	ref, err := fresh.deserializeNodeWithHash(rootBlob.Blob, 0, before)
+	if err != nil {
+		t.Fatalf("deserializeNodeWithHash: %v", err)
+	}
+	if ref.Kind() != kindInternal {
+		t.Fatalf("expected root to be Internal, got kind %d", ref.Kind())
+	}
+	// Spot-check: the reload-tree's root hash equals the commit-time hash.
+	if got := fresh.computeHash(ref); got != before {
+		t.Fatalf("reload root hash mismatch: got %x, want %x", got, before)
+	}
+}
+
+// TestDecodeRejectsNonCanonicalPosition hand-crafts a blob where the bitmap
+// position has nonzero trailing bits given its depth offset. Two
+// implementations must produce byte-identical blobs for the same logical
+// content, so a non-canonical position is unambiguously an invalid blob.
+func TestDecodeRejectsNonCanonicalPosition(t *testing.T) {
+	// groupDepth=5, bitmap size = 4 bytes. Set bit at position 5 (binary
+	// 00101) and declare depthOffset=2. Top 2 bits of 00101 are 00 (path
+	// "00"), the trailing 3 bits should be zero — they're 101 here, so the
+	// reader must reject.
+	blob := []byte{nodeTypeInternal, 5}
+	// bitmap[0] = bit at position 5 → 1 << (7-5) = 0x04
+	blob = append(blob, 0x04, 0x00, 0x00, 0x00)
+	// depths[0] = 2, packed as (2-1)=1 in 3 bits MSB-first → 0b0010_0000 = 0x20
+	blob = append(blob, 0x20)
+	// hashes[0] = 32 zero bytes
+	blob = append(blob, make([]byte, HashSize)...)
+
+	s := newNodeStore()
+	_, err := s.deserializeNode(blob, 0)
+	if err == nil {
+		t.Fatal("expected non-canonical position error, got nil")
+	}
+	if err.Error() != "non-canonical bitmap position" {
+		t.Errorf("expected 'non-canonical bitmap position', got %q", err.Error())
+	}
+}
+
+// TestDecodeRejectsInvalidDepthOffset covers depthOffset>groupDepth (the entry
+// would live below the group's bottom layer, impossible by construction). The
+// old depthOffset=0 and depthOffset>MaxGroupDepth cases are gone: the 3-bit
+// field stores (offset-1) ∈ [0,7], so offset 0 and offset 9 are unrepresentable
+// and can no longer be hand-crafted into a blob. Only offset>groupDepth with
+// groupDepth<MaxGroupDepth remains encodable and must still be rejected.
+func TestDecodeRejectsInvalidDepthOffset(t *testing.T) {
+	makeBlob := func(groupDepth int, depthOffset uint8) []byte {
+		bitmapSize := bitmapSizeForDepth(groupDepth)
+		bitmap := make([]byte, bitmapSize)
+		bitmap[0] = 0x80 // bit at position 0
+		depths := make([]byte, packedDepthsLen(1))
+		writeDepth(depths, 0, depthOffset-1)
+		blob := []byte{nodeTypeInternal, byte(groupDepth)}
+		blob = append(blob, bitmap...)
+		blob = append(blob, depths...)
+		blob = append(blob, make([]byte, HashSize)...)
+		return blob
+	}
+
+	for _, tc := range []struct {
+		name        string
+		groupDepth  int
+		depthOffset uint8
+	}{
+		{"gd2/depth3", 2, 3},
+		{"gd3/depth4", 3, 4},
+		{"gd5/depth6", 5, 6},
+		{"gd7/depth8", 7, 8},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			s := newNodeStore()
+			_, err := s.deserializeNode(makeBlob(tc.groupDepth, tc.depthOffset), 0)
+			if err == nil {
+				t.Fatal("expected invalid depth offset error, got nil")
+			}
+			if err.Error() != "invalid depth offset" {
+				t.Errorf("expected 'invalid depth offset', got %q", err.Error())
+			}
+		})
+	}
+}
+
+// TestDecodeRejectsNonCanonicalDepthPadding verifies the canonical-encoding
+// check on the packed depth stream: when k*3 is not a multiple of 8, the unused
+// low bits of the final packed byte must be zero. Here groupDepth=2 with a
+// single entry uses 3 bits, leaving 5 pad bits; a stray pad bit must be
+// rejected so that two encoders cannot produce differing blobs for the same
+// content.
+func TestDecodeRejectsNonCanonicalDepthPadding(t *testing.T) {
+	// groupDepth=2, one entry at bitmap position 0, depthOffset=2 (bottom layer).
+	// Packed depth = (2-1)=1 → 0b001_00000 = 0x20; set a stray low pad bit → 0x21.
+	blob := []byte{nodeTypeInternal, 2}
+	blob = append(blob, 0x80) // bitmap: bit at position 0
+	blob = append(blob, 0x21) // packed depths with a non-zero pad bit
+	blob = append(blob, make([]byte, HashSize)...)
+
+	s := newNodeStore()
+	_, err := s.deserializeNode(blob, 0)
+	if err == nil {
+		t.Fatal("expected non-canonical depth padding error, got nil")
+	}
+	if err.Error() != "non-canonical depth padding" {
+		t.Errorf("expected 'non-canonical depth padding', got %q", err.Error())
+	}
+}
+
+// TestRoundTripPersistence exercises the full commit-then-reload pipeline
+// the way real geth does it: write blobs to a backing map, open a fresh
+// nodeStore from the root blob, then resolve and read every value back
+// through the resolver. Catches mismatches between the writer's storage
+// path (collectChildGroups in store_commit.go) and the reader's lookup
+// path (keyToPath in store_ops.go) — exactly the bug Option B's first
+// implementation hit, where blobs were written at the bottom-layer-extended
+// path but resolved at the natural-depth path. Also confirms the reloaded
+// trie's root hash equals the original committed hash.
+func TestRoundTripPersistence(t *testing.T) {
+	for _, groupDepth := range []int{1, 2, 3, 5, 8} {
+		t.Run(fmt.Sprintf("groupDepth=%d", groupDepth), func(t *testing.T) {
+			// 1. Build a trie with deterministically-distributed keys.
+			//    50 keys with FNV-style spread guarantees several stems
+			//    land in the root group (natural depth ≤ groupDepth) and
+			//    several land in sub-groups, exercising both the in-group
+			//    resolve and the cross-group resolve paths.
+			writerTrie := &BinaryTrie{
+				store:      newNodeStore(),
+				tracer:     trie.NewPrevalueTracer(),
+				groupDepth: groupDepth,
+			}
+			const n = 50
+			keys := make([][HashSize]byte, n)
+			values := make([][HashSize]byte, n)
+			for i := range n {
+				binary.BigEndian.PutUint64(keys[i][:8], uint64(i+1)*0x9e3779b97f4a7c15)
+				binary.BigEndian.PutUint64(keys[i][8:16], uint64(i+1)*0xc2b2ae3d27d4eb4f)
+				binary.BigEndian.PutUint64(keys[i][16:24], uint64(i+1)*0x165667b19e3779f9)
+				binary.BigEndian.PutUint64(keys[i][24:32], uint64(i+1)*0x85ebca77c2b2ae63)
+				binary.BigEndian.PutUint64(values[i][:8], uint64(i+1))
+				if err := writerTrie.store.Insert(keys[i][:], values[i][:], nil); err != nil {
+					t.Fatalf("insert %d: %v", i, err)
+				}
+			}
+
+			// 2. Commit; capture every blob into an in-memory map keyed
+			//    by its path. The NodeSet key is the BitArray.PutKeyBytes
+			//    encoding — exactly the bytes the resolver gets from
+			//    keyToPath, so map lookups by string(path) round-trip.
+			rootHash := writerTrie.Hash()
+			_, ns := writerTrie.Commit(false)
+			blobs := make(map[string][]byte, len(ns.Nodes))
+			for path, node := range ns.Nodes {
+				blobs[path] = node.Blob
+			}
+
+			// 3. Build a resolver that serves blobs from the map.
+			resolver := func(path []byte, hash common.Hash) ([]byte, error) {
+				blob, ok := blobs[string(path)]
+				if !ok {
+					return nil, fmt.Errorf("blob not found at path %x (hash %x)", path, hash)
+				}
+				return blob, nil
+			}
+
+			// 4. Open a fresh store, seeded only with the root blob.
+			//    Everything else must be reached via the resolver.
+			readerStore := newNodeStore()
+			rootBlob, ok := blobs[""]
+			if !ok {
+				t.Fatalf("no root blob in NodeSet")
+			}
+			rootRef, err := readerStore.deserializeNodeWithHash(rootBlob, 0, rootHash)
+			if err != nil {
+				t.Fatalf("deserialize root: %v", err)
+			}
+			readerStore.root = rootRef
+
+			// 5. Read every key back through the resolver and verify.
+			//    A mismatch here means either the storage path diverged
+			//    from the lookup path, or deserialization corrupted data.
+			for i := range n {
+				got, err := readerStore.Get(keys[i][:], resolver)
+				if err != nil {
+					t.Fatalf("Get key %d (%x): %v", i, keys[i], err)
+				}
+				if !bytes.Equal(got, values[i][:]) {
+					t.Fatalf("Get key %d: got %x, want %x", i, got, values[i][:])
+				}
+			}
+
+			// 6. The reloaded trie's root hash must equal the original.
+			//    This is the canonical-hash round-trip property: any
+			//    independent reader walking the same blobs computes the
+			//    same root, independent of in-memory layout choices.
+			if got := readerStore.Hash(); got != rootHash {
+				t.Fatalf("post-reload root hash: got %x, want %x", got, rootHash)
+			}
+		})
+	}
+}
+
+// TestNoOrphanBlobAfterStemPromotion targets gballet's store_ops.go review
+// concern: when a second commit promotes an existing stem deeper, the stem's
+// blob moves to a new path, and Commit emits only AddNode entries (never
+// deletes). If a stem's old path were not reoccupied by the new ancestor node,
+// the prior commit's blob would linger as an unreachable orphan.
+//
+// The test applies two commit deltas to a single backing map, then walks the
+// trie from the new root and asserts every persisted blob is reachable — i.e.
+// no orphan survives. The first batch establishes stems at group boundaries;
+// the second batch shares prefixes with the first to force promotions.
+func TestNoOrphanBlobAfterStemPromotion(t *testing.T) {
+	for _, groupDepth := range []int{1, 2, 3, 5} {
+		t.Run(fmt.Sprintf("groupDepth=%d", groupDepth), func(t *testing.T) {
+			tr := &BinaryTrie{
+				store:      newNodeStore(),
+				tracer:     trie.NewPrevalueTracer(),
+				groupDepth: groupDepth,
+			}
+			db := make(map[string][]byte)
+			apply := func(ns *trienode.NodeSet) {
+				for path, node := range ns.Nodes {
+					if node.IsDeleted() {
+						delete(db, path)
+						continue
+					}
+					db[path] = node.Blob
+				}
+			}
+
+			const n = 24
+			keys := make([][HashSize]byte, n)
+			values := make([][HashSize]byte, n)
+			for i := range n {
+				binary.BigEndian.PutUint64(keys[i][:8], uint64(i+1)*0x9e3779b97f4a7c15)
+				binary.BigEndian.PutUint64(keys[i][8:16], uint64(i+1)*0xc2b2ae3d27d4eb4f)
+				binary.BigEndian.PutUint64(values[i][:8], uint64(i+1))
+			}
+
+			// Commit 1: first half.
+			for i := 0; i < n/2; i++ {
+				if err := tr.store.Insert(keys[i][:], values[i][:], nil); err != nil {
+					t.Fatalf("insert %d: %v", i, err)
+				}
+			}
+			_, ns1 := tr.Commit(false)
+			apply(ns1)
+
+			// Commit 2: second half (shares prefixes, forces promotions).
+			for i := n / 2; i < n; i++ {
+				if err := tr.store.Insert(keys[i][:], values[i][:], nil); err != nil {
+					t.Fatalf("insert %d: %v", i, err)
+				}
+			}
+			rootHash, ns2 := tr.Commit(false)
+			apply(ns2)
+
+			// Walk from the new root, recording every blob the reader resolves.
+			resolved := make(map[string]bool)
+			resolver := func(path []byte, _ common.Hash) ([]byte, error) {
+				resolved[string(path)] = true
+				blob, ok := db[string(path)]
+				if !ok {
+					return nil, fmt.Errorf("missing blob at path %x", path)
+				}
+				return blob, nil
+			}
+			reader := newNodeStore()
+			rootRef, err := reader.deserializeNodeWithHash(db[""], 0, rootHash)
+			if err != nil {
+				t.Fatalf("deserialize root: %v", err)
+			}
+			reader.root = rootRef
+
+			for i := range n {
+				got, err := reader.Get(keys[i][:], resolver)
+				if err != nil {
+					t.Fatalf("Get key %d: %v", i, err)
+				}
+				if !bytes.Equal(got, values[i][:]) {
+					t.Fatalf("Get key %d: got %x, want %x", i, got, values[i][:])
+				}
+			}
+
+			// Every persisted blob must be reachable; the root ("") is seeded.
+			reachable := map[string]bool{"": true}
+			for path := range resolved {
+				reachable[path] = true
+			}
+			for path := range db {
+				if !reachable[path] {
+					t.Errorf("orphan blob at path %x is unreachable from the new root", []byte(path))
+				}
+			}
+		})
+	}
+}
+
+// stemsDivergingAt returns two 32-byte stems whose first `divergeBit` bits
+// are zero and whose bit at index `divergeBit` differs (left=0, right=1).
+// Useful for placing two stems at a known natural depth within a group.
+func stemsDivergingAt(divergeBit int) (left, right []byte) {
+	left = make([]byte, HashSize)
+	right = make([]byte, HashSize)
+	// Bit `divergeBit` is in byte (divergeBit/8) at MSB position (7 - divergeBit%8).
+	right[divergeBit/8] = 1 << (7 - divergeBit%8)
+	return left, right
+}
+
+// bitFlipStem returns a 32-byte stem whose first `divergeBit` bits are zero,
+// bit `divergeBit` is 1, and all subsequent bits are zero. Used together
+// with the all-zero stem to force divergence at a specific bit.
+func bitFlipStem(divergeBit int) []byte {
+	out := make([]byte, HashSize)
+	out[divergeBit/8] = 1 << (7 - divergeBit%8)
+	return out
+}
--- a/trie/bintrie/internal_node.go
+++ b/trie/bintrie/internal_node.go
@ -26,12 +26,8 @@ func keyToPath(depth int, key []byte) ([]byte, error) {
 	if depth >= 31*8 {
 		return nil, errors.New("node too deep")
 	}
-	path := make([]byte, 0, depth+1)
-	for i := range depth + 1 {
-		bit := key[i/8] >> (7 - (i % 8)) & 1
-		path = append(path, bit)
-	}
-	return path, nil
+	path := new(BitArray).SetBytes(uint8(depth+1), key)
+	return path.KeyBytes(), nil
 }

 // Invariant: dirty=false implies mustRecompute=false. Every mutation that
--- a/trie/bintrie/internal_node_test.go
+++ b/trie/bintrie/internal_node_test.go
@ -283,14 +283,13 @@ func TestInternalNodeCollectNodes(t *testing.T) {
 		t.Fatal(err)
 	}

-	var collectedPaths [][]byte
-	flushFn := func(path []byte, hash common.Hash, serialized []byte) {
-		pathCopy := make([]byte, len(path))
-		copy(pathCopy, path)
-		collectedPaths = append(collectedPaths, pathCopy)
+	var collectedPaths []BitArray
+	flushFn := func(path BitArray, hash common.Hash, serialized []byte) {
+		collectedPaths = append(collectedPaths, path)
 	}

-	s.collectNodes(s.root, []byte{1}, flushFn, 8)
+	initialPath := NewBitArray(1, 1)
+	s.collectNodes(s.root, initialPath, flushFn, 8)

 	// Should have collected 3 nodes: left stem, right stem, and the internal node itself
 	if len(collectedPaths) != 3 {
--- a/trie/bintrie/iterator.go
+++ b/trie/bintrie/iterator.go
@ -188,20 +188,20 @@ func (it *binaryNodeIterator) Parent() common.Hash {
 	return it.store.computeHash(it.stack[len(it.stack)-2].Node)
 }

-// Path returns the bit-path to the current node.
+// Path returns the bit-packed path to the current node.
 // Callers must not retain references to the returned slice after calling Next.
 func (it *binaryNodeIterator) Path() []byte {
 	if it.Leaf() {
 		return it.LeafKey()
 	}
-	var path []byte
+	var path BitArray
 	for i, state := range it.stack {
 		if i >= len(it.stack)-1 {
 			break
 		}
-		path = append(path, byte(state.Index))
+		path.AppendBit(&path, uint8(state.Index))
 	}
-	return path
+	return path.KeyBytes()
 }

 func (it *binaryNodeIterator) NodeBlob() []byte {
--- a/trie/bintrie/node_store.go
+++ b/trie/bintrie/node_store.go
@ -41,10 +41,15 @@ type nodeStore struct {
 	// stem-split keeps the old stem at a deeper position), so they don't
 	// have free lists.
 	freeHashed []uint32
+
+	// orphans holds on-disk paths whose committed blob has been abandoned by a
+	// stem depth-promotion since the last commit. Commit emits a deletion for
+	// each (unless a freshly flushed node reoccupies the path), then clears it.
+	orphans map[string]struct{}
 }

 func newNodeStore() *nodeStore {
-	return &nodeStore{root: emptyRef}
+	return &nodeStore{root: emptyRef, orphans: make(map[string]struct{})}
 }

 func (s *nodeStore) allocInternal() uint32 {
@ -179,6 +184,10 @@ func (s *nodeStore) Copy() *nodeStore {
 		ns.freeHashed = make([]uint32, len(s.freeHashed))
 		copy(ns.freeHashed, s.freeHashed)
 	}
+	ns.orphans = make(map[string]struct{}, len(s.orphans))
+	for path := range s.orphans {
+		ns.orphans[path] = struct{}{}
+	}

 	return ns
 }
--- a/trie/bintrie/stem_node_test.go
+++ b/trie/bintrie/stem_node_test.go
@ -313,14 +313,13 @@ func TestStemNodeCollectNodes(t *testing.T) {
 		t.Fatal(err)
 	}

-	var collectedPaths [][]byte
-	flushFn := func(path []byte, hash common.Hash, serialized []byte) {
-		pathCopy := make([]byte, len(path))
-		copy(pathCopy, path)
-		collectedPaths = append(collectedPaths, pathCopy)
+	var collectedPaths []BitArray
+	flushFn := func(path BitArray, hash common.Hash, serialized []byte) {
+		collectedPaths = append(collectedPaths, path)
 	}

-	s.collectNodes(s.root, []byte{0, 1, 0}, flushFn, 8)
+	initialPath := NewBitArray(3, 0b010)
+	s.collectNodes(s.root, initialPath, flushFn, 8)

 	// Should have collected one node (itself)
 	if len(collectedPaths) != 1 {
@ -328,7 +327,8 @@ func TestStemNodeCollectNodes(t *testing.T) {
 	}

 	// Check the path
-	if !bytes.Equal(collectedPaths[0], []byte{0, 1, 0}) {
-		t.Errorf("Path mismatch: expected [0, 1, 0], got %v", collectedPaths[0])
+	expectedPath := NewBitArray(3, 0b010)
+	if !collectedPaths[0].Equal(&expectedPath) {
+		t.Errorf("Path mismatch: expected %v, got %v", expectedPath, collectedPaths[0])
 	}
 }
--- a/trie/bintrie/store_commit.go
+++ b/trie/bintrie/store_commit.go
@ -27,7 +27,7 @@ import (
 	"github.com/ethereum/go-ethereum/common"
 )

-type nodeFlushFn func(path []byte, hash common.Hash, serialized []byte)
+type nodeFlushFn func(path BitArray, hash common.Hash, serialized []byte)

 func (s *nodeStore) Hash() common.Hash {
 	return s.computeHash(s.root)
@ -111,7 +111,7 @@ func (s *nodeStore) hashInternal(idx uint32) common.Hash {
 // It traverses up to `remainingDepth` levels, storing hashes of bottom-layer children.
 // position tracks the current index (0 to 2^groupDepth - 1) for bitmap placement.
 // hashes collects the hashes of present children, bitmap tracks which positions are present.
-func (s *nodeStore) serializeSubtree(ref nodeRef, remainingDepth int, position int, absoluteDepth int, bitmap []byte, hashes *[]common.Hash) {
+func (s *nodeStore) serializeSubtree(ref nodeRef, remainingDepth int, position int, groupDepth int, bitmap []byte, hashes *[]common.Hash, depths *[]uint8) {
 	if remainingDepth == 0 {
 		// Bottom layer: store hash if not empty
 		switch ref.Kind() {
@ -122,6 +122,7 @@ func (s *nodeStore) serializeSubtree(ref nodeRef, remainingDepth int, position i
 			// StemNode, HashedNode, or InternalNode at boundary: store hash
 			bitmap[position/8] |= 1 << (7 - (position % 8))
 			*hashes = append(*hashes, s.computeHash(ref))
+			*depths = append(*depths, uint8(groupDepth))
 		}
 		return
 	}
@ -130,57 +131,81 @@ func (s *nodeStore) serializeSubtree(ref nodeRef, remainingDepth int, position i
 	case kindInternal:
 		leftPos := position * 2
 		rightPos := position*2 + 1
-		s.serializeSubtree(s.getInternal(ref.Index()).left, remainingDepth-1, leftPos, absoluteDepth+1, bitmap, hashes)
-		s.serializeSubtree(s.getInternal(ref.Index()).right, remainingDepth-1, rightPos, absoluteDepth+1, bitmap, hashes)
+		s.serializeSubtree(s.getInternal(ref.Index()).left, remainingDepth-1, leftPos, groupDepth, bitmap, hashes, depths)
+		s.serializeSubtree(s.getInternal(ref.Index()).right, remainingDepth-1, rightPos, groupDepth, bitmap, hashes, depths)
 	case kindEmpty:
 		return
 	default:
 		// StemNode or HashedNode encountered before reaching the group's bottom
 		// layer. Compute the leaf bitmap position where this node's hash will
 		// be stored.
-		leafPos := position
-		switch ref.Kind() {
-		case kindStem:
-			sn := s.getStem(ref.Index())
-			// Extend position using the stem's key bits so that
-			// GetValuesAtStem traversal (which follows key bits) finds the hash.
-			for d := 0; d < remainingDepth; d++ {
-				bit := sn.Stem[(absoluteDepth+d)/8] >> (7 - ((absoluteDepth + d) % 8)) & 1
-				leafPos = leafPos*2 + int(bit)
-			}
-		default:
-			// HashedNode or unknown: extend all-left (no key bits available).
-			// This matches the all-zero path that resolveNode would follow.
-			leafPos = position << remainingDepth
-		}
-		bitmap[leafPos/8] |= 1 << (7 - (leafPos % 8))
+		bitmapPos := position << remainingDepth
+		bitmap[bitmapPos/8] |= 1 << (7 - (bitmapPos % 8))
 		*hashes = append(*hashes, s.computeHash(ref))
+		*depths = append(*depths, uint8(groupDepth-remainingDepth))
 	}
 }

+// depthBits is the number of bits used to encode one depth offset.
+const depthBits = 3
+
+// packedDepthsLen returns the byte length of k packed depth entries
+func packedDepthsLen(k int) int {
+	return (k*depthBits + 7) / 8
+}
+
+// writeDepth writes a depth entry at idx into the buf, MSB-first.
+func writeDepth(buf []byte, idx int, v uint8) {
+	pos := idx * depthBits
+	for i := range depthBits {
+		bit := (v >> (depthBits - 1 - i)) & 1
+		p := pos + i
+		buf[p>>3] |= bit << (7 - (p & 7))
+	}
+}
+
+// readDepth reads a depth for entry idx from buf.
+func readDepth(buf []byte, idx int) uint8 {
+	pos := idx * depthBits
+	var v uint8
+	for i := range depthBits {
+		p := pos + i
+		bit := (buf[p>>3] >> (7 - (p & 7))) & 1
+		v = v<<1 | bit
+	}
+	return v
+}
+
 // SerializeNode serializes a node into the flat on-disk format.
 func (s *nodeStore) serializeNode(ref nodeRef, groupDepth int) []byte {
 	switch ref.Kind() {
 	case kindInternal:
-		// InternalNode group: 1 byte type + 1 byte group depth + variable bitmap + N×32 byte hashes
+		// InternalNode group format:
+		//   [type(1)] [groupDepth(1)] [bitmap (2^groupDepth bits)] [depths(3 bits × K, padded)] [hashes(32B × K)]
 		bitmapSize := bitmapSizeForDepth(groupDepth)
 		bitmap := make([]byte, bitmapSize)
 		var hashes []common.Hash
+		var depths []uint8

-		node := s.getInternal(ref.Index())
-		s.serializeSubtree(ref, groupDepth, 0, int(node.depth), bitmap, &hashes)
+		s.serializeSubtree(ref, groupDepth, 0, groupDepth, bitmap, &hashes, &depths)

 		// Build serialized output
-		serializedLen := NodeTypeBytes + 1 + bitmapSize + len(hashes)*HashSize
+		k := len(hashes)
+		depthsLen := packedDepthsLen(k)
+		serializedLen := NodeTypeBytes + 1 + bitmapSize + depthsLen + k*HashSize
 		serialized := make([]byte, serializedLen)
 		serialized[0] = nodeTypeInternal
-		serialized[1] = byte(groupDepth) // group depth => bitmap size for a sparse group
+		serialized[1] = byte(groupDepth)
 		copy(serialized[2:2+bitmapSize], bitmap)

-		offset := NodeTypeBytes + 1 + bitmapSize
-		for _, h := range hashes {
-			copy(serialized[offset:offset+HashSize], h.Bytes())
-			offset += HashSize
+		depthsOff := NodeTypeBytes + 1 + bitmapSize
+		for i, d := range depths {
+			writeDepth(serialized[depthsOff:depthsOff+depthsLen], i, d-1)
+		}
+
+		hashesOff := depthsOff + depthsLen
+		for i, h := range hashes {
+			copy(serialized[hashesOff+i*HashSize:hashesOff+(i+1)*HashSize], h.Bytes())
 		}

 		return serialized
@ -229,56 +254,90 @@ func (s *nodeStore) deserializeNodeWithHash(serialized []byte, depth int, hn com
 }

 // deserializeSubtree reconstructs an InternalNode subtree from grouped serialization.
-// remainingDepth is how many more levels to build, position is current index in the bitmap,
-// nodeDepth is the actual trie depth for the node being created.
-// hashIdx tracks the current position in the hash data (incremented as hashes are consumed).
-func (s *nodeStore) deserializeSubtree(hn common.Hash, remainingDepth int, position int, nodeDepth int, bitmap []byte, hashData []byte, hashIdx *int, mustRecompute bool, dirty bool) (nodeRef, error) {
-	if remainingDepth == 0 {
-		// Bottom layer: check bitmap and return HashedNode or Empty
-		if bitmap[position/8]>>(7-(position%8))&1 == 1 {
-			if len(hashData) < (*hashIdx+1)*HashSize {
-				return emptyRef, errInvalidSerializedLength
-			}
-			hash := common.BytesToHash(hashData[*hashIdx*HashSize : (*hashIdx+1)*HashSize])
-			*hashIdx++
-			return s.newHashedRef(hash), nil
-		}
+func (s *nodeStore) deserializeSubtree(hn common.Hash, groupDepth int, nodeDepth int, bitmap []byte, depths []byte, hashData []byte, mustRecompute bool, dirty bool) (nodeRef, error) {
+	if len(hashData)%HashSize != 0 {
+		return emptyRef, errInvalidSerializedLength
+	}
+	k := len(hashData) / HashSize
+	if len(depths) != packedDepthsLen(k) {
+		return emptyRef, errInvalidSerializedLength
+	}
+	if k == 0 {
 		return emptyRef, nil
 	}

-	// Check if this entire subtree is empty by examining all relevant bitmap bits
-	leftPos := position * 2
-	rightPos := position*2 + 1
-
-	// note that the parent might not need root computations, but the children
-	// do, because their hash isn't saved. Hence `mustRecompute` is set to `true`.
-	left, err := s.deserializeSubtree(common.Hash{}, remainingDepth-1, leftPos, nodeDepth+1, bitmap, hashData, hashIdx, true, dirty)
-	if err != nil {
-		return emptyRef, err
-	}
-	right, err := s.deserializeSubtree(common.Hash{}, remainingDepth-1, rightPos, nodeDepth+1, bitmap, hashData, hashIdx, true, dirty)
-	if err != nil {
-		return emptyRef, err
-	}
-
-	// If both children are empty, return Empty
-	if left.IsEmpty() && right.IsEmpty() {
-		return emptyRef, nil
-	}
-
-	ref := s.newInternalRef(nodeDepth)
-	node := s.getInternal(ref.Index())
-	node.left = left
-	node.right = right
-	node.mustRecompute = mustRecompute
+	rootRef := s.newInternalRef(nodeDepth)
+	rootNode := s.getInternal(rootRef.Index())
+	rootNode.mustRecompute = mustRecompute
 	if !mustRecompute {
-		// mustRecompute will only be false for the root of the subtree,
-		// for which we already know the hash.
-		node.hash = hn
-		node.mustRecompute = false
+		rootNode.hash = hn
 	}
-	node.dirty = dirty
-	return ref, nil
+	rootNode.dirty = dirty
+
+	bitmapBits := 1 << groupDepth
+	entryIdx := 0
+	for bit := 0; bit < bitmapBits; bit++ {
+		if bitmap[bit/8]>>(7-(bit%8))&1 == 0 {
+			continue
+		}
+		depthOffset := int(readDepth(depths, entryIdx)) + 1
+		if depthOffset > groupDepth {
+			return emptyRef, errors.New("invalid depth offset")
+		}
+		// Canonical-encoding check: trailing position bits must be zero.
+		mask := (1 << (groupDepth - depthOffset)) - 1
+		if bit&mask != 0 {
+			return emptyRef, errors.New("non-canonical bitmap position")
+		}
+		var hash common.Hash
+		copy(hash[:], hashData[entryIdx*HashSize:(entryIdx+1)*HashSize])
+		if err := s.attachInGroup(rootRef, nodeDepth, groupDepth, depthOffset, bit, hash, dirty); err != nil {
+			return emptyRef, err
+		}
+		entryIdx++
+	}
+	return rootRef, nil
+}
+
+func (s *nodeStore) attachInGroup(rootRef nodeRef, rootDepth, groupDepth, depthOffset, bitmapPos int, hash common.Hash, dirty bool) error {
+	cur := rootRef
+	for level := 0; level < depthOffset-1; level++ {
+		bit := (bitmapPos >> (groupDepth - 1 - level)) & 1
+		node := s.getInternal(cur.Index())
+		childRef := node.left
+		if bit == 1 {
+			childRef = node.right
+		}
+		if childRef.IsEmpty() {
+			newRef := s.newInternalRef(rootDepth + level + 1)
+			s.getInternal(newRef.Index()).dirty = dirty
+			if bit == 0 {
+				node.left = newRef
+			} else {
+				node.right = newRef
+			}
+			cur = newRef
+			continue
+		}
+		if childRef.Kind() != kindInternal {
+			return errors.New("overlapping entries in group blob")
+		}
+		cur = childRef
+	}
+	leafBit := (bitmapPos >> (groupDepth - depthOffset)) & 1
+	node := s.getInternal(cur.Index())
+	if leafBit == 0 {
+		if !node.left.IsEmpty() {
+			return errors.New("overlapping entries in group blob")
+		}
+		node.left = s.newHashedRef(hash)
+	} else {
+		if !node.right.IsEmpty() {
+			return errors.New("overlapping entries in group blob")
+		}
+		node.right = s.newHashedRef(hash)
+	}
+	return nil
 }

 func (s *nodeStore) decodeNode(serialized []byte, depth int, hn common.Hash, mustRecompute, dirty bool) (nodeRef, error) {
@ -288,7 +347,9 @@ func (s *nodeStore) decodeNode(serialized []byte, depth int, hn common.Hash, mus

 	switch serialized[0] {
 	case nodeTypeInternal:
-		// Grouped format: 1 byte type + 1 byte group depth + variable bitmap + N×32 byte hashes
+		// Grouped format:
+		//   [type(1)] [groupDepth(1)] [bitmap (2^groupDepth bits, padded to bitmapSize bytes)]
+		//   [depthOffsets (3 bits × K, padded to bytes)] [hashes (32B × K)]
 		if len(serialized) < NodeTypeBytes+1 {
 			return emptyRef, errInvalidSerializedLength
 		}
@ -301,10 +362,38 @@ func (s *nodeStore) decodeNode(serialized []byte, depth int, hn common.Hash, mus
 			return 0, errInvalidSerializedLength
 		}
 		bitmap := serialized[2 : 2+bitmapSize]
-		hashData := serialized[2+bitmapSize:]

-		hashIdx := 0
-		return s.deserializeSubtree(hn, groupDepth, 0, depth, bitmap, hashData, &hashIdx, mustRecompute, dirty)
+		bitmapBits := 1 << groupDepth
+		if bitmapBits < 8 {
+			padMask := byte(0xFF) >> bitmapBits
+			if bitmap[0]&padMask != 0 {
+				return emptyRef, errors.New("non-canonical bitmap padding")
+			}
+		}
+
+		k := 0
+		for _, b := range bitmap {
+			k += bits.OnesCount8(b)
+		}
+		depthsLen := packedDepthsLen(k)
+		expectedLen := NodeTypeBytes + 1 + bitmapSize + depthsLen + k*HashSize
+		if len(serialized) != expectedLen {
+			return emptyRef, errInvalidSerializedLength
+		}
+		depthsOff := NodeTypeBytes + 1 + bitmapSize
+		depths := serialized[depthsOff : depthsOff+depthsLen]
+		hashData := serialized[depthsOff+depthsLen : depthsOff+depthsLen+k*HashSize]
+
+		// Canonical-encoding check: the unused low bits of the last packed
+		// depth byte must be zero.
+		if usedBits := k * depthBits; usedBits%8 != 0 {
+			padMask := byte(0xFF) >> (usedBits % 8)
+			if depths[depthsLen-1]&padMask != 0 {
+				return emptyRef, errors.New("non-canonical depth padding")
+			}
+		}
+
+		return s.deserializeSubtree(hn, groupDepth, depth, bitmap, depths, hashData, mustRecompute, dirty)

 	case nodeTypeStem:
 		if len(serialized) < NodeTypeBytes+StemSize+StemBitmapSize {
@ -340,7 +429,10 @@ func (s *nodeStore) decodeNode(serialized []byte, depth int, hn common.Hash, mus
 // CollectNodes flushes every node that needs flushing via flushfn in post-order.
 // Invariant: any ancestor of a node that needs flushing is itself marked, so a
 // clean root means the whole subtree is clean.
-func (s *nodeStore) collectNodes(ref nodeRef, path []byte, flushfn nodeFlushFn, groupDepth int) {
+//
+// BitArray is passed by value (33 bytes) to keep child paths on the stack.
+// Passing by pointer causes escape to heap per recursive call.
+func (s *nodeStore) collectNodes(ref nodeRef, path BitArray, flushfn nodeFlushFn, groupDepth int) {
 	switch ref.Kind() {
 	case kindInternal:
 		node := s.getInternal(ref.Index())
@ -375,77 +467,51 @@ func (s *nodeStore) collectNodes(ref nodeRef, path []byte, flushfn nodeFlushFn,
 // collectChildGroups traverses within a group to find and collect nodes in the next group.
 // remainingLevels is how many more levels below the current node until we reach the group boundary.
 // When remainingLevels=0, the current node's children are at the next group boundary.
-func (s *nodeStore) collectChildGroups(node *InternalNode, path []byte, flushfn nodeFlushFn, groupDepth int, remainingLevels int) error {
+func (s *nodeStore) collectChildGroups(node *InternalNode, path BitArray, flushfn nodeFlushFn, groupDepth int, remainingLevels int) error {
 	if remainingLevels == 0 {
 		// Current node is at depth (groupBoundary - 1), its children are at the next group boundary
 		if !node.left.IsEmpty() {
-			s.collectNodes(node.left, appendBit(path, 0), flushfn, groupDepth)
+			leftPath := path
+			leftPath.AppendBit(&leftPath, 0)
+			s.collectNodes(node.left, leftPath, flushfn, groupDepth)
 		}
 		if !node.right.IsEmpty() {
-			s.collectNodes(node.right, appendBit(path, 1), flushfn, groupDepth)
+			rightPath := path
+			rightPath.AppendBit(&rightPath, 1)
+			s.collectNodes(node.right, rightPath, flushfn, groupDepth)
 		}
 		return nil
 	}

 	if !node.left.IsEmpty() {
+		leftPath := path
+		leftPath.AppendBit(&leftPath, 0)
 		switch node.left.Kind() {
 		case kindInternal:
 			n := s.getInternal(node.left.Index())
-			if err := s.collectChildGroups(n, appendBit(path, 0), flushfn, groupDepth, remainingLevels-1); err != nil {
+			if err := s.collectChildGroups(n, leftPath, flushfn, groupDepth, remainingLevels-1); err != nil {
 				return err
 			}
 		default:
-			extPath := s.extendPathToGroupLeaf(appendBit(path, 0), node.left, remainingLevels)
-			s.collectNodes(node.left, extPath, flushfn, groupDepth)
+			s.collectNodes(node.left, leftPath, flushfn, groupDepth)
 		}
 	}
 	if !node.right.IsEmpty() {
+		rightPath := path
+		rightPath.AppendBit(&rightPath, 1)
 		switch node.right.Kind() {
 		case kindInternal:
 			n := s.getInternal(node.right.Index())
-			if err := s.collectChildGroups(n, appendBit(path, 1), flushfn, groupDepth, remainingLevels-1); err != nil {
+			if err := s.collectChildGroups(n, rightPath, flushfn, groupDepth, remainingLevels-1); err != nil {
 				return err
 			}
 		default:
-			extPath := s.extendPathToGroupLeaf(appendBit(path, 1), node.right, remainingLevels)
-			s.collectNodes(node.right, extPath, flushfn, groupDepth)
+			s.collectNodes(node.right, rightPath, flushfn, groupDepth)
 		}
 	}
 	return nil
 }

-// extendPathToGroupLeaf extends a storage path to the group's leaf boundary,
-// matching the projection done by serializeSubtree. For StemNodes, the path
-// is extended using the stem's key bits (same as serializeSubtree). For other
-// node types, the path is extended with all-zero (left) bits.
-func (s *nodeStore) extendPathToGroupLeaf(path []byte, node nodeRef, remainingLevels int) []byte {
-	if remainingLevels <= 0 {
-		return path
-	}
-	if node.Kind() == kindStem {
-		sn := s.getStem(node.Index())
-		for _ = range remainingLevels {
-			bit := sn.Stem[len(path)/8] >> (7 - (len(path) % 8)) & 1
-			path = appendBit(path, bit)
-		}
-	} else {
-		// HashedNode or other: all-left extension (matches serializeSubtree's
-		// position << remainingDepth behavior).
-		for _ = range remainingLevels {
-			path = appendBit(path, 0)
-		}
-	}
-	return path
-}
-
-// appendBit appends a bit to a path, returning a new slice
-func appendBit(path []byte, bit byte) []byte {
-	var p [256]byte
-	copy(p[:], path)
-	result := p[:len(path)]
-	return append(result, bit)
-}
-
 func (s *nodeStore) toDot(ref nodeRef, parent, path string) string {
 	switch ref.Kind() {
 	case kindInternal:
--- a/trie/bintrie/store_ops.go
+++ b/trie/bintrie/store_ops.go
@ -262,7 +262,15 @@ func (s *nodeStore) splitStemValuesInsert(existingRef nodeRef, newStem []byte, v
 	bitStem := existing.Stem[existing.depth/8] >> (7 - (existing.depth % 8)) & 1
 	nRef := s.newInternalRef(int(existing.depth))
 	nNode := s.getInternal(nRef.Index())
+	if !existing.dirty {
+		var buf [33]byte
+		oldPath := new(BitArray).SetBytes(existing.depth, existing.Stem[:]).PutKeyBytes(buf[:])
+		s.orphans[string(oldPath)] = struct{}{}
+	}
 	existing.depth++
+	// The existing stem's on-disk path lengthens by one bit, which means
+	// the stem must be re-flushed at the longer new path.
+	existing.dirty = true

 	bitKey := newStem[nNode.depth/8] >> (7 - (nNode.depth % 8)) & 1
 	if bitKey == bitStem {
--- a/trie/bintrie/trie.go
+++ b/trie/bintrie/trie.go
@ -347,12 +347,35 @@ func (t *BinaryTrie) Hash() common.Hash {
 func (t *BinaryTrie) Commit(_ bool) (common.Hash, *trienode.NodeSet) {
 	nodeset := trienode.NewNodeSet(common.Hash{})

-	// Pre-size the path buffer: collectNodes reuses it in-place via
-	// append/truncate; 32 covers typical binary-trie depth without regrowth.
-	pathBuf := make([]byte, 0, 32)
-	t.store.collectNodes(t.store.root, pathBuf, func(path []byte, hash common.Hash, serialized []byte) {
-		nodeset.AddNode(path, trienode.NewNodeWithPrev(hash, serialized, t.tracer.Get(path)))
+	// Stem depth-promotion abandons a committed blob and is the only source of
+	// orphans. When none are pending (the common case) we skip tracking flushed
+	// paths entirely, keeping Commit allocation-free beyond the node set.
+	var added map[string]struct{}
+	if len(t.store.orphans) > 0 {
+		added = make(map[string]struct{})
+	}
+	var rootPath BitArray
+	t.store.collectNodes(t.store.root, rootPath, func(path BitArray, hash common.Hash, serialized []byte) {
+		var buf [33]byte
+		pathBytes := path.PutKeyBytes(buf[:])
+		if added != nil {
+			added[string(pathBytes)] = struct{}{}
+		}
+		nodeset.AddNode(pathBytes, trienode.NewNodeWithPrev(hash, serialized, t.tracer.Get(pathBytes)))
 	}, t.groupDepth)
+
+	// Delete blobs abandoned by stem depth-promotion, unless a freshly flushed
+	// node already reoccupies the path (the group-boundary case).
+	if len(t.store.orphans) > 0 {
+		for path := range t.store.orphans {
+			if _, ok := added[path]; ok {
+				continue
+			}
+			nodeset.AddNode([]byte(path), trienode.NewDeletedWithPrev(t.tracer.Get([]byte(path))))
+		}
+		t.store.orphans = make(map[string]struct{})
+	}
+
 	return t.Hash(), nodeset
 }