Durable episode integrity

This feature applies to RocksDbMemoryStorage, used by PersistentAgentMemory. The existing HTTP memory routes use AgentMemory and remain volatile; this change does not make those routes persistent.

Guarantees#

• An episode's embedded agent ID must match its storage scope. IDs cannot be reassigned to another agent. Wrong-agent reads return no result; wrong-agent deletion returns false without touching the owner’s index.
• Moving an episode's event time atomically deletes its old row and updates its row and UUID index in a single RocksDB write batch.
• A per-storage mutex serializes index read/modify/write operations. Cloned Arc handles share this lock; RocksDB excludes a second independent opener. This does not provide optimistic conflict detection for application-level read/modify/write sequences.
• All mutation paths honor WAL and synchronous-write settings. With sync_writes=true and enable_wal=true, acknowledged writes request WAL synchronization. Disabling WAL reduces durability. Explicit flush covers all memory column families.
• Index parsing checks lengths before accessing bytes and returns a corruption error on malformed records.

Record format and compatibility#

New episode values have the six-byte QMEP\0\x01 header followed by a bincode tuple (Episode, Option<String>). The episode's structured data field is temporarily empty inside the binary component; the second component contains its JSON representation. This avoids bincode's lack of deserialize_any support, required by serde_json::Value, without changing episode keys or binary property values.

Unversioned bincode episodes without structured data remain readable and are upgraded on subsequent writes. Unknown envelope versions fail explicitly. Legacy records already written with structured JSON could not be decoded by the original implementation and are not repaired by this feature. Downgrading to an older binary after writing new records is unsupported; back up before upgrading. No existing database is rewritten automatically.

Validation#

Eight regression tests in storage.rs cover cross-agent delete protection, owner mismatch and UUID reassignment, timestamp moves, structured JSON across reopen, malformed indexes, legacy reads/updates, concurrent moves, invalid scopes and unknown versions. Five original reproductions failed before the fix. Validation commands:

cargo test --workspace --all-targets --locked
rustfmt --edition 2024 --check crates/qilbee-memory/src/storage.rs

Reopen tests exercise normal process shutdown and recovery, not power-loss or filesystem fault injection. The existing HTTP authorization, graph transaction atomicity and application-level access-counter races require separate work.