The native state, history, request, metadata, balance, block and export record domains share the existing Database transaction with typed shards and ApplicationStore content, history, request receipts and claims. These tests exercise the public RecordDomain/RecordLimits API added by the native acceptance integration; the owning implementation and test-module wiring are maintained in that integration change.
Validation on macOS arm64, Rust 1.95.0, using
CARGO_TARGET_DIR=/tmp/bbg-migration-target and --features backend-ssd,backend-hdd:
cargo test --manifest-path rs/Cargo.toml --locked --features backend-ssd,backend-hdd --test native_records: 3 tests pass, each running against both Fjall and redb.cargo test --manifest-path rs/Cargo.toml --locked --features backend-ssd,backend-hdd --lib storage::database::native_tests: 4 tests pass.- Scoped rustfmt/whitespace checks pass. Clippy for the library and native_records integration target succeeds with four existing BBG warnings and three existing vendor warnings; none originate in the added integration tests.
The integration cases verify all seven domains remain isolated despite sharing keys; one successful application transaction updates/deletes native records and shards alongside its receipt/content/claims; a rejected closure rolls every domain back; committed changes and markers agree after reopen; identical retries skip the transition; cloned application and shard views retain the exclusive writer lock until the last owner drops. Native point reads, prefix scans, exclusive cursors, byte budgets and staged-table scan rejection are exercised.
The fault tests wrap a real redb FileBackend with a one-shot failure and assert the backend consumed it. They cover a physical read after closure staging but before commit, a write failure, failure before synchronization, and failure returning from a synchronization that already completed. Unknown commit outcomes freeze native record reads/scans, transactions, application views and shard views. Reopen checks all native records, native deletions, application history/content/ receipt, typed shards and commit markers as one old-or-new outcome. A completed barrier with a failed acknowledgement recovers the new state, and retry does not execute the accepted transition twice.
An observed redb behavior matters for recovery: a precommit read I/O failure seals redb's handle and subsequent reads explicitly require reopen. That is a known-aborted closure rather than CommitUnknown; the test reopens before checking rollback. It does not assume that every known I/O failure leaves a usable handle.
These cases establish storage-component composition. They do not execute a Cybergraph transition validator, simulate physical power loss, or qualify the complete node acceptance path. Native coordination and node-process tests are owned by the dependent repositories.