diff --git a/packages/cultnet-rs/tests/byte_slot_migration.rs b/packages/cultnet-rs/tests/byte_slot_migration.rs
new file mode 100644
index 00000000..0acdcd4d
--- /dev/null
+++ b/packages/cultnet-rs/tests/byte_slot_migration.rs
@@ -0,0 +1,113 @@
+//! Does adding `#[cultcache(key = N, bytes)]` to a field break reading the
+//! records already written without it?
+//!
+//! The attribute changes a slot from a MessagePack array-of-integers (serde's
+//! default for `Vec<u8>`) to a real `bin`. Six of cultnet-rs's byte fields
+//! already carry it and four do not, so converting the remaining four is a
+//! stored-data question before it is a signature question: every record
+//! written before the flip holds an array in that slot.
+//!
+//! Both directions are tested, because a migration has to survive a rollback
+//! as well as a roll-forward.
+
+use cultcache_rs::DatabaseEntry;
+use serde::Serialize;
+
+#[derive(Clone, Debug, PartialEq, Eq, DatabaseEntry)]
+#[cultcache(type = "test.byte_slot", schema = "test.byte_slot.v1")]
+struct ArrayEncoded {
+    #[cultcache(key = 0)]
+    schema_version: String,
+    #[cultcache(key = 1)]
+    payload: Vec<u8>,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, DatabaseEntry)]
+#[cultcache(type = "test.byte_slot", schema = "test.byte_slot.v1")]
+struct BinEncoded {
+    #[cultcache(key = 0)]
+    schema_version: String,
+    #[cultcache(key = 1, bytes)]
+    payload: Vec<u8>,
+}
+
+fn sample() -> Vec<u8> {
+    (0u8..=255).collect()
+}
+
+#[test]
+fn the_attribute_is_what_switches_the_wire_shape() {
+    let array = rmp_serde::to_vec(&ArrayEncoded {
+        schema_version: "test.byte_slot.v1".into(),
+        payload: sample(),
+    })
+    .expect("array encodes");
+    let bin = rmp_serde::to_vec(&BinEncoded {
+        schema_version: "test.byte_slot.v1".into(),
+        payload: sample(),
+    })
+    .expect("bin encodes");
+
+    // 256 bytes as `bin` is a 3-byte header plus the payload. As an array it
+    // is one or two bytes per element, because every value above 0x7f needs a
+    // u8 marker of its own.
+    println!("array-encoded {} bytes, bin-encoded {} bytes", array.len(), bin.len());
+    assert!(
+        array.len() > bin.len(),
+        "array {} should exceed bin {}",
+        array.len(),
+        bin.len()
+    );
+}
+
+#[test]
+fn a_bytes_reader_still_reads_records_written_before_the_attribute() {
+    let written_before = rmp_serde::to_vec(&ArrayEncoded {
+        schema_version: "test.byte_slot.v1".into(),
+        payload: sample(),
+    })
+    .expect("array encodes");
+
+    // serde_bytes::ByteBuf accepts an array marker as well as bin, so the
+    // flip is NOT a hard read break. This was worth measuring rather than
+    // assuming: the opposite would have forced a dual-read path or a rewrite
+    // migration onto every field with persisted records.
+    let read_after =
+        rmp_serde::from_slice::<BinEncoded>(&written_before).expect("bytes reader reads an array");
+    assert_eq!(read_after.payload, sample());
+}
+
+#[test]
+fn a_pre_attribute_reader_still_reads_records_written_after_it() {
+    let written_after = rmp_serde::to_vec(&BinEncoded {
+        schema_version: "test.byte_slot.v1".into(),
+        payload: sample(),
+    })
+    .expect("bin encodes");
+
+    // The rollback direction, so reverting the binary is survivable.
+    let read_by_old =
+        rmp_serde::from_slice::<ArrayEncoded>(&written_after).expect("plain reader reads bin");
+    assert_eq!(read_by_old.payload, sample());
+}
+
+#[test]
+fn re_encoding_an_old_record_under_the_attribute_does_not_reproduce_its_bytes() {
+    // Where the cost actually lands. Reading is compatible; reproducing is
+    // not. Idunn byte-exact-checks its control store by re-encoding what it
+    // decoded, and canonical_sha256 hashes the same serialization -- so a
+    // record written before the flip can be read, but can never re-encode to
+    // the bytes it was stored as, and its canonical digest changes.
+    let written_before = rmp_serde::to_vec(&ArrayEncoded {
+        schema_version: "test.byte_slot.v1".into(),
+        payload: sample(),
+    })
+    .expect("array encodes");
+    let decoded = rmp_serde::from_slice::<BinEncoded>(&written_before).expect("reads");
+    let re_encoded = rmp_serde::to_vec(&decoded).expect("re-encodes");
+
+    assert_ne!(
+        written_before, re_encoded,
+        "identical bytes would mean the attribute changed nothing"
+    );
+}