Testkit (MOD-TESTKIT)¶
Seeded, portable input generation and oracles for every kernel test (PRD 05 §7, 12). Everything is deterministic from a uint64 seed and byte-identical across x86-64, ARM64, and macOS — enforced in CI by committed golden FNV-1a hashes (Testkit.GoldenHashesAreCrossPlatformStable).
Components¶
| File | Contents |
|---|---|
tests/testkit/generators.h/.cpp |
SplitMix64 Rng; QLM-1 axis implementations (PRD 11 §4): sequential/uniform/Zipf(θ=1.0, 1000 values) value distributions, uniform/clustered/alternating bitmap patterns, selection-vector derivation, boundary-value and float-special sets, AlignedBuffer offset factory |
tests/testkit/reference.h |
The naive second oracle (dual-oracle scheme, REQ-TEST-002): LSB-first bit accessors, reference popcount, ADR-016 tail checks. Family oracles land with their families (M3/M6) |
tests/testkit/assertions.h |
First-divergence diagnostics: index, ±4-element hex context, seed, requirement ID (REQ-TEST-012 format) |
tests/testkit/drift_check.cpp |
Conformance runner asserting the bench-local distribution copies are byte-identical (REQ-BENCH-015); intentionally free of GoogleTest and Google Benchmark (REQ-TEST-018) |
Portability rules (REQ-INT-002)¶
Integer-only core; uniform doubles built as (next() >> 11) * 2⁻⁵³; Zipf CDF from harmonic sums (IEEE +// only — no libm anywhere); geometric run lengths via Bernoulli trials. These rules are what make the golden hashes platform-independent.
Updating the generator spec¶
The bench harness maintains an independent copy of the same spec (bench/harness/distributions.*, REQ-BENCH-015). A change therefore touches, in one PR: both implementations, the golden hashes (regenerate with quiver_drift_check --print-golden), and a CHANGELOG note. Testkit.BenchDistributionDriftAlarm (ctest) fails if the copies diverge.
Reproduction recipe¶
Every failure message carries the seed and axis values (REQ-TEST-012); re-run with --gtest_filter=<Test> — generation is deterministic, so the failure reproduces exactly.