I use this standalone experiment to compare the two complete Metal merge paths on identical logical inputs. It does not change the library or the existing GPU calibration. The retained matched results include all twelve cases. Both formats merge two sorted, unique streams with no cancellations. The integration record identifies a public commit whose complete measured source closure matches the original hashes.

Fixtures

Each key is four unsigned 32-bit lanes. KV03 receives those lanes as sixteen big-endian bytes, preserving the fixed format's ordering. Both paths receive exactly the same present value bytes. The fixed format omits the optional-value and sort framing required by KV03; that difference is part of the formats being compared.

The twelve cases combine structured or hash-like keys with:

FF: sixteen-byte values, 4,096 / 65,536 / 262,144 records per input.
FV: values of zero through 512 bytes, 4,096 / 65,536 / 131,072 records per input.

The structured keys come directly from the fixed experiment's shared fixture header. The hash-like keys apply a bijective 64-bit mixer to each logical ID for the high half and a second mixed half, then sort outside timing. Uniqueness follows from the high-half bijection. This is a deterministic hash-like distribution, not a cryptographic hash workload. Canonical logical input/output files record all key and value bytes, and the collector checks their SHA-256 equality across both formats and every fresh process.

The existing KV parser's conservative maximum-record-size times record-count bound excludes the largest FV case at 262,144 records per input. I retain that guard and stop at 131,072 for both formats.

Measurement boundary

The comparable complete_ms timer includes fresh output creation/mapping, Metal imports of existing mapped inputs, temporary allocation, the complete GPU merge, CPU constant metadata work, body and header CRC32C, file clipping, and return-time resource cleanup. Fixed output CRCs use previously unused header words 19 and 20; the header checksum sees its own field as zero. This wrapper adds integrity work without changing the fixed merge kernels or their layout. KV03 retains its normal checksums. Neither path calls fsync or msync, publishes a catalog root, or constructs a fractional index.

Fixture generation, input encoding/mapping, CPU correctness oracles, shader and pipeline loading, and full output verification are outside timing. One excluded warmup precedes three measured iterations in every fresh process. Every output must equal the corresponding CPU encoder's entire canonical file, including checksums. The CPU oracle is also checked against the shared logical records.

The original internal timers are retained as diagnostic columns. Their scopes are different: the fixed timer starts with imported inputs and output, while the KV timer includes its own fresh output mapping and checksum. Only the outer complete timers are the matched comparison. Fixed checksum time is separately available; KV checksum time remains included in its complete timer.

The KV configuration is the previously calibrated optimized default: mapped compressed inputs, GPU frame parsing and prefix-owner reconstruction, eight-byte prefix cache, word emitter, and GPU output EF. The tree remains levelwise and the output plan separate. The measured tiled-tree and fused-plan experiments remain opt-in after mixed complete-path results. This comparison does not claim to find the fastest possible KV implementation. Fixed uses its existing no-cancellation fast path and Merge Path 32 ordering.

Reproduction

Use the C++26 module toolchain, Ninja and an installed simd package with exceptions enabled. CMake builds the host and links Everett from this source checkout; Python handles shader artifacts and manifests.

First build the two existing optional packages and supply their compiled Metal libraries. Shader sources must match the retained artifact provenance. The wrappers include the existing host implementations verbatim and compile only new program entry points; they do not rebuild or modify the shaders.

EVERETT_EXPERIMENT_ARCH=AUTO selects NEON on an ARM host with the native Everett archive, and scalar elsewhere. Select AVX2 or AVX512 explicitly only on a matching x86 build host; configuration checks its required features. SCALAR is always available. The experiment's policies carry that architecture explicitly, and new collections record it in their build metadata. Existing retained measurements describe their original sources and compiler settings.

cmake -G Ninja -DCMAKE_PREFIX_PATH=/path/to/simd -S optional/fixed_kv_compare -B build-fixed-kv \
  -DEVERETT_KV_METALLIB=/path/to/kv/kernels.metallib \
  -DEVERETT_FIXED_METALLIB=/path/to/fixed/kernels.metallib
cmake --build build-fixed-kv --parallel 1
ctest --test-dir build-fixed-kv --output-on-failure --parallel 1
python3 optional/fixed_kv_compare/collect.py \
  --build build-fixed-kv --output /path/to/new-results

Use the host GPU resource gate around the serial checks and collection. Commit a clean source checkpoint before collecting. The fixed schedule uses three paired fresh processes per case, alternating format order across rounds (a necessary 2:1 first-slot split with three rounds), and rotates case order by four each round. No result-dependent reruns or case selection are part of the collector.