This suite compares the actual v0.5.0 native runner at df7f61300fbb9222f20022127ec0058cc0c9c8d1 with v0.6.0. It is a format-loading benchmark, not an optimizer-quality, complete-job or storage-throughput claim. Raw evidence and CSV metadata stay under ignored build/ directories.

The historical v0.6.0 result measured format loading only. Since v0.7.0, row-only TPE parents reconstruct without proposal replay; exact continuation requires the optional sampler checkpoint. The no-replay suite measures import through the first new trial, tests checkpoint continuation against the old replay oracle, and compares exact full-history versus opted-in bounded batch cost.

Resource protocol

Use an otherwise idle little-endian Linux machine with 8 vCPU. Each measured process runs in its own cgroup with MemoryMax=8G and MemorySwapMax=0; this includes file-backed pages charged to that cgroup. The acceptance host has 16 GiB physically, with this enforced 8-GiB process budget. GNU time records process RSS and elapsed time. Generation and file hashing are outside the measured boundary.

python3 benchmarks/warm/run_benchmark.py \
  --baseline build/baseline/bin/pineforge-hpo-native \
  --native build/bin/pineforge-hpo-native \
  --output build/warm-resources

The driver generates exact unique deterministic permuted histories at 100k/500k/1M/2M rows and 5/32 inputs. Five-input histories use float64 columns; 32-input histories use four int32 grids plus 28 float64 columns. All objectives are finite with deliberate ties. The equivalent JSON includes three report metrics. One cold-page-cache run (drop_caches) and three subsequent warm-cache runs are retained for each format/size. It requires sudo for transient cgroups and cache eviction. Failed loads retain their actual error/time/RSS; they are never presented as successful load timings. v0.5.0 refuses documents above 256 MiB, independently of available memory.

The measured command is native space-info --spec ... --warm-start ...: file read/mmap, all scalar/schema/hash validation and an exact tried-vector index. It does not load a strategy/data, replay historical TPE proposals, fit density models or render ancestor JSON. Binary payload SHA-256 is deferred; the unmodified v0.5 JSON loader includes its original digest computation. Ingest/download/object-hash work normally precedes this loader and populates the OS page cache. Cold times also measure the provisioned disk and can exceed the subsecond warm-load target.

resources.csv.metadata.json includes the fixture seed, memory budget, versions, baseline revision, compiler flags, executable SHA-256, every generated object hash, CSV hash and measured boundary. Preserve it alongside resources.csv. For a copied build tree without Git metadata, pass --source-revision FULL_COMMIT to bind the measurement to the reviewed source revision.

Bitwise continuation gate

python3 tests/test_warm_binary_contract.py \
  build/bin/pineforge-hpo-native build/lib/fake_strategy.so \
  --baseline build/baseline/bin/pineforge-hpo-native \
  --output build/warm-equivalence

The gate compares packed uint64 IDs and binary64 parameter/objective bits, states and typed categorical values, not approximate float equality. It covers grid, random, TPE and constrained TPE, finite/mixed/log spaces, and batches 1/4/8/32. Each v0.5 parent feeds v0.5 JSON, v0.6 JSON, one binary block and reversed chunks. The native converter and Python writer must match bytes. The JSON child of a binary parent must preserve its full ancestry for a grandchild. Evidence includes per-case suggestion and warm-object SHA-256 in equivalence.json.

The native golden/corruption tests validate every truncation point, versions, counts, flags, descriptors, states, indices, duplicate IDs and wide IDs. Python goldens validate the same writer bytes. The real-strategy gate also runs binary and multi-block continuations against JSON, with canonical harness metric/trade comparison:

python3 tests/test_warm_start_e2e.py \

--native build/bin/pineforge-hpo-native --output build/warm-real-strategy

Run Release, ASan/UBSan and TSan builds as documented in the root build options, the full Python suite, the pinned Optuna smoke profile in benchmarks/README.md, and the zero-warning Doxygen/generated-site gates before release claims.