diff --git a/src/cachekit/cache_handler.py b/src/cachekit/cache_handler.py
index 0f47dfcd..0a924ec9 100644
--- a/src/cachekit/cache_handler.py
+++ b/src/cachekit/cache_handler.py
@@ -1354,7 +1354,7 @@ async def _handle_l2_read_error_async(self, e: SerializationError, cache_key: st
             get_logger().warning(f"Failed to evict poisoned L2 entry {redact_cache_key(cache_key)}: {del_err}")
         self._notify_deserialize_error(e, cache_key)
 
-    def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) -> Optional[Any]:
+    def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) -> Optional[tuple[bool, Any, int]]:
         """Get value from cache if it exists.
 
         Args:
@@ -1362,7 +1362,11 @@ def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) ->
             refresh_ttl: Optional TTL to refresh on hit
 
         Returns:
-            Tuple (True, value) if cache hit, None if cache miss or error
+            Tuple (True, value, size_bytes) if cache hit, None if cache miss or error.
+            size_bytes is the length of the L2 envelope actually served, the same quantity
+            the L1 hit site records via len(l1_bytes) (LAB-3768). It is a length rather than
+            the envelope itself: the sync path does no L1 backfill (#164) and the mmap fast
+            path's view dangles once its handle closes (#171).
 
         Note:
             Requires cache_handler to be set via set_cache_handler() before calling.
@@ -1381,7 +1385,9 @@ def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) ->
                 if handle is not None:
                     try:
                         get_logger().cache_hit(cache_key, "Backend(mmap)")
-                        return (True, self.serialization_handler.deserialize_data(handle.view, cache_key))
+                        value = self.serialization_handler.deserialize_data(handle.view, cache_key)
+                        # Size is read before close(); the view itself must never leave this frame (#171).
+                        return (True, value, handle.view.nbytes)
                     finally:
                         handle.close()
 
@@ -1390,8 +1396,8 @@ def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) ->
                 get_logger().cache_hit(cache_key, "Backend")
                 # Pass cache_key for AAD verification (required for encrypted data)
                 deserialized = self.serialization_handler.deserialize_data(cached_data, cache_key)
-                # Return a tuple (True, value) to distinguish from "no cache entry"
-                return (True, deserialized)
+                # Tuple distinguishes a hit from "no cache entry"; the served envelope size rides along
+                return (True, deserialized, len(cached_data))
             return None
         except KeyringConfigurationError:
             # LOCAL keyring config fault (bad tenant_id, bad keyring entry index) —
@@ -1410,11 +1416,11 @@ def get_cached_value(self, cache_key: str, refresh_ttl: Optional[int] = None) ->
             get_logger().warning(f"Backend operation failed for get on {cache_key}: {e}")
             return None
 
-    def get_cached_value_with_freshness(self, cache_key: str) -> Optional[tuple[tuple[bool, Any], bool, Optional[int]]]:
+    def get_cached_value_with_freshness(self, cache_key: str) -> Optional[tuple[tuple[bool, Any, int], bool, Optional[int]]]:
         """SWR variant of :meth:`get_cached_value` (LAB-381/LAB-557): also reports
         staleness and the server's remaining freshness in seconds.
 
-        Returns ``((True, value), is_stale, fresh_for)`` on a hit, None on
+        Returns ``((True, value, size_bytes), is_stale, fresh_for)`` on a hit, None on
         miss/error. fresh_for is None when no signal exists (pre-signal server,
         non-SWR backend) — the caller applies legacy L1 TTL behavior. The mmap
         fast path is skipped — SWR is CachekitIO-only, which is not buffer-readable.
@@ -1436,7 +1442,7 @@ def get_cached_value_with_freshness(self, cache_key: str) -> Optional[tuple[tupl
             cached_data, is_stale, fresh_for = hit
             get_logger().cache_hit(cache_key, "Backend(stale)" if is_stale else "Backend")
             deserialized = self.serialization_handler.deserialize_data(cached_data, cache_key)
-            return ((True, deserialized), is_stale, fresh_for)
+            return ((True, deserialized, len(cached_data)), is_stale, fresh_for)
         except KeyringConfigurationError:
             # LOCAL keyring config fault (bad tenant_id, bad keyring entry index) —
             # never a legitimate miss, and not tamper. Re-raised past the broad
diff --git a/src/cachekit/decorators/wrapper.py b/src/cachekit/decorators/wrapper.py
index 08b2209b..c3c8af0f 100644
--- a/src/cachekit/decorators/wrapper.py
+++ b/src/cachekit/decorators/wrapper.py
@@ -1322,7 +1322,8 @@ def sync_wrapper(*args: Any, **kwargs: Any) -> Any:  # noqa: PLR0912
             duration = time.time() - start_time
 
             if cached_result is not None:
-                # Cached result is a tuple (True, actual_value)
+                # (True, value, size_bytes): served L2 envelope length, see get_cached_value.
+                _found, result, size_bytes = cached_result
                 features.set_operation_context("get", duration_ms=duration * 1000)
                 features.record_success()
 
@@ -1337,7 +1338,6 @@ def sync_wrapper(*args: Any, **kwargs: Any) -> Any:  # noqa: PLR0912
                     )
 
                 # Record cache hit with structured logging
-                size_bytes = len(str(cached_result[1]).encode("utf-8")) if cached_result[1] is not None else 0
                 features.log_cache_operation(
                     operation="get",
                     key=cache_key,
@@ -1350,7 +1350,6 @@ def sync_wrapper(*args: Any, **kwargs: Any) -> Any:  # noqa: PLR0912
 
                 # Also record statistics if enabled
                 if features.collect_stats:
-                    size_bytes = len(str(cached_result[1]).encode("utf-8")) if cached_result[1] is not None else 0
                     features.record_cache_operation(
                         operation="get",
                         namespace=namespace or "default",
@@ -1374,7 +1373,7 @@ def sync_wrapper(*args: Any, **kwargs: Any) -> Any:  # noqa: PLR0912
                 # WHY: L2 cache hit returns from try block that lacks finally cleanup
                 # (only inner try at line ~567, not the outer try-finally at ~645-720)
                 reset_current_function_stats(token)
-                return cached_result[1]
+                return result
         except DecryptionAuthenticationError:
             # Fail-closed tamper failure propagated from get_cached_value — it only
             # raises when encryption.fail_closed=True (the metric and error log were
diff --git a/tests/unit/test_decrypt_fail_policy.py b/tests/unit/test_decrypt_fail_policy.py
index 3961708d..c913443d 100644
--- a/tests/unit/test_decrypt_fail_policy.py
+++ b/tests/unit/test_decrypt_fail_policy.py
@@ -332,7 +332,7 @@ def test_fail_open_valid_entry_roundtrips(self):
         """Regression: the happy path is untouched by the policy plumbing."""
         handler, strategy, serialization = _make_operation_handler(fail_closed=True)
         strategy.store["key:a"] = serialization.serialize_data({"v": 7}, cache_key="key:a")
-        assert handler.get_cached_value("key:a") == (True, {"v": 7})
+        assert handler.get_cached_value("key:a") == (True, {"v": 7}, len(strategy.store["key:a"]))
 
 
 class TestConfigDriftRead:
diff --git a/tests/unit/test_mmap_read_path.py b/tests/unit/test_mmap_read_path.py
index 8f08731d..c01e941e 100644
--- a/tests/unit/test_mmap_read_path.py
+++ b/tests/unit/test_mmap_read_path.py
@@ -99,12 +99,13 @@ def test_eligible_reads_via_mmap_and_confines_the_handle(self) -> None:
         sh.supports_mmap_read.return_value = True
         sh.deserialize_data.return_value = sentinel
         handle = MagicMock()
+        handle.view = memoryview(b"payload")
         ch = MagicMock()
         ch.get_buffer.return_value = handle
 
         result = self._handler(sh, ch).get_cached_value("k")
 
-        assert result == (True, sentinel)
+        assert result == (True, sentinel, len(b"payload"))  # size is the mapped payload, read before close
         ch.get_buffer.assert_called_once_with("k")
         ch.get.assert_not_called()  # normal read path NOT used on the mmap hit
         sh.deserialize_data.assert_called_once_with(handle.view, "k")
@@ -133,7 +134,7 @@ def test_get_buffer_none_falls_through_to_normal_read(self) -> None:
 
         ch.get_buffer.assert_called_once()
         ch.get.assert_called_once()  # fell through
-        assert result == (True, "val")
+        assert result == (True, "val", len(b"frame"))
 
 
 @pytest.mark.unit
@@ -156,7 +157,8 @@ def test_arrow_dataframe_roundtrips_through_real_mmap(self, tmp_path) -> None:
         oh = CacheOperationHandler(sh, CacheKeyGenerator(), cache_handler=ch)
 
         df = pd.DataFrame({"a": range(2000), "b": [float(i) / 3 for i in range(2000)]})
-        ch.set("k", sh.serialize_data(df, cache_key="k"), 300)
+        envelope = sh.serialize_data(df, cache_key="k")
+        ch.set("k", envelope, 300)
 
         with (
             patch.object(backend, "get", wraps=backend.get) as g,
@@ -165,8 +167,9 @@ def test_arrow_dataframe_roundtrips_through_real_mmap(self, tmp_path) -> None:
             hit = oh.get_cached_value("k")
 
         assert hit is not None
-        found, value = hit
+        found, value, size_bytes = hit
         assert found is True
         pd.testing.assert_frame_equal(value, df)
+        assert size_bytes == len(envelope)  # the mmap hit reports the served envelope, not a repr estimate (LAB-3768)
         gb.assert_called_once()  # the real mmap path was taken
         g.assert_not_called()  # not the os.read fallback
diff --git a/tests/unit/test_swr_decorator.py b/tests/unit/test_swr_decorator.py
index ffe6dc36..a93ce8ba 100644
--- a/tests/unit/test_swr_decorator.py
+++ b/tests/unit/test_swr_decorator.py
@@ -528,7 +528,7 @@ def test_legacy_two_tuple_handler_degrades_to_no_bound(self) -> None:
         cache_handler = mock.MagicMock()
         cache_handler.get_with_freshness.return_value = (b"bytes", False)  # 0.5.x 2-tuple
         op.set_cache_handler(cache_handler)
-        assert op.get_cached_value_with_freshness("k") == ((True, {"v": 1}), False, None)
+        assert op.get_cached_value_with_freshness("k") == ((True, {"v": 1}, len(b"bytes")), False, None)
 
     def test_backend_error_reads_as_miss(self) -> None:
         op, cache_handler = self._make_op()
diff --git a/tests/unit/test_sync_l2_hit_size_bytes.py b/tests/unit/test_sync_l2_hit_size_bytes.py
new file mode 100644
index 00000000..b80a7be0
--- /dev/null
+++ b/tests/unit/test_sync_l2_hit_size_bytes.py
@@ -0,0 +1,62 @@
+"""Sync L2-hit ``size_bytes`` measures the served envelope (LAB-3768).
+
+The sync L1 hit site records ``len(l1_bytes)``; the sync L2 site recorded
+``len(str(value).encode())``, a repr estimate of the deserialized value, because the sync
+handler returned no envelope. A secure cache (ciphertext envelope vs plaintext repr) got a
+bimodal histogram on one series.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from cachekit import cache
+from cachekit.decorators.orchestrator import FeatureOrchestrator
+
+
+class _ByteStore:
+    """Plain in-memory byte store standing in for L2."""
+
+    def __init__(self) -> None:
+        self.store: dict[str, bytes] = {}
+
+    def get(self, key: str) -> bytes | None:
+        return self.store.get(key)
+
+    def set(self, key: str, value: bytes, ttl: int | None = None) -> None:
+        self.store[key] = value
+
+    def delete(self, key: str) -> bool:
+        return self.store.pop(key, None) is not None
+
+    def exists(self, key: str) -> bool:
+        return key in self.store
+
+    def health_check(self) -> tuple[bool, dict[str, Any]]:
+        return True, {}
+
+
+@pytest.mark.unit
+def test_sync_l2_hit_records_envelope_size(monkeypatch: pytest.MonkeyPatch) -> None:
+    # Patched at the orchestrator, not the collector: record_success() also forwards an
+    # operation-context record to the collector, which would shadow the value under test.
+    recorded: list[dict[str, Any]] = []
+    monkeypatch.setattr(FeatureOrchestrator, "record_cache_operation", lambda self, **kw: recorded.append(kw))
+    backend = _ByteStore()
+
+    @cache(backend=backend, l1_enabled=False, ttl=60, namespace="sync-l2-size")
+    def compute() -> dict[str, int]:
+        return {"answer": 42}
+
+    assert compute() == {"answer": 42}  # miss: primes L2
+    (envelope,) = backend.store.values()
+    recorded.clear()
+
+    assert compute() == {"answer": 42}
+
+    gets = [c for c in recorded if c["operation"] == "get"]
+    assert len(gets) == 1
+    assert gets[0]["size_bytes"] == len(envelope)
+    assert len(envelope) != len(str({"answer": 42}).encode())  # non-vacuous: the old repr estimate differs