Tombstones carried no ciphertext, so the server could forge a delete (hiding
an item), and a device that pulled one accepted the next live version without
a seq comparison (delete-then-replay rollback).
* A delete is sealed like an item (same AAD and padding) with plaintext
{id, seq, deleted: true}; seq = working/base seq + 1. DELETE carries it as
the ItemPut body; the test server stores and serves it.
* replay: Base is None | LegacyTombstone | Sealed(plaintext). check_pulled
refuses a record without ciphertext, a deleted flag that disagrees with the
plaintext, or a tombstone naming another item, then requires seq to grow
for tombstones too. The `restores` exception is gone: edit beats delete
re-seals with max(edit, tombstone) + 1. next_version returns just seq.
* Legacy: a pulled tombstone without ciphertext is an integrity alarm (cannot
be told apart from a forgery). A pending legacy delete is sealed on push; a
legacy tombstone already held as base is accepted as unknown seq.
* engine: open_sealed, seal_tombstone; next_local_version -> u64. Merkle leaf
hashes a sealed tombstone's envelope like a live record's.
* kv-model Item: `restores` removed. kv-ffi delete_item seals the tombstone;
FFI surface unchanged.
* Tests: forged (unsealed and relabelled) tombstones raise an integrity alarm
(kv-sync and kv-ffi), replay after a tombstone, replays of the tombstone and
of the restore, legacy pending delete, shared item_replay.json table; the
live test now deletes an item through the real server.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
samiandClaude Opus 5.5
cfab3ab543