Validation & testing

Local validation

The recorded validation on October 16, 2026 covered Windows amd64 with Go 1.26.0 and Linux amd64 under Ubuntu WSL2. Windows tests, vet, a CLI build, and the runnable example passed. Linux vet and CGO_ENABLED=1 go test -race ./... also passed.

The Windows environment did not have a C compiler for local race execution. The GitHub workflow defines compiler-equipped Windows and Linux jobs; workflow definitions are not a substitute for inspecting a particular CI run.

Correctness coverage

• CRUD, namespaces, byte ownership, empty values, versions, and restart behavior.
• Atomic batches, repeated keys, failed validation, concurrent consistent scans.
• Range merges, tombstones, budgets, automatic flush, and reclamation.
• Every incomplete byte offset of a final two-operation WAL frame.
• Checksum failures, whole-record-boundary table truncation, and missing storage.
• Checkpoint crash images and reclamation of a full old checkpointed WAL.
• Forced child-process termination after durable acknowledgement.
• A deterministic 500-step reference-model workload with flushes and reopenings.
• Poisoned-handle behavior after real WAL I/O failure.

Run checks

From the repository root:

go vet ./...
go test ./...
go run ./examples/basic

With a supported C compiler and CGO enabled:

go test -race ./...

Fuzz the frame and record decoders:

go test ./internal/storage -run='^$' -fuzz=FuzzDecode -fuzztime=30s
go test ./internal/storage -run='^$' -fuzz=FuzzReadFrame -fuzztime=30s

The initial short decoder and frame-reader smoke runs executed 1,508,110 and 118,376 cases respectively, without a discovered failure. Short fuzz runs are not exhaustive proof.

Benchmark interpretation

go test -run='^$' -bench=. -benchmem -count=3 .

The current benchmarks repeatedly read or replace one 128-byte hot-key value. Reads distinguish memory and warm-cache table access; writes distinguish durable and relaxed mode.

Results report averages and Go allocations. They do not establish P50/P95/P99, RSS, cold-cache behavior, application-scale throughput, or fair cross-database superiority. See the raw baseline and methodology.

Remaining hardening

Sudden power loss, device faults, long-running stress, network filesystems, audited backup/restore, migrations, and macOS/FreeBSD runtime behavior have not been validated. Checkpoint crash images model protocol boundaries rather than every filesystem instruction.

The complete historical record is kept in docs/validation.md.