The integrated implementation reduces median complete-merge time by 12.50–80.31% across the twelve M2 Max fixtures. Zero-added-prefix byte keys improve by 12.50–17.11%, resolving the earlier short-byte median regression in this paired run. The four 4 KiB-prefix cases improve by 63.12–80.31%.

These are direct measurements of e45e47f14b45a2f9726d23c80d1d19b33de7745a against 4429df0. The final candidate carries exact longest-common-prefix (LCP) lengths between merge heads, handles a successor's first policy unit directly, and shares internal frame output without retaining a third full key in the merger. It also includes the generic Elias–Fano/profile refactor, cold exceptions, explicit unit shifts and additive byte rounding. This combined comparison does not assign its gains to any single change.

Final direct comparison

Median nanoseconds per distinct output key; negative change means less time. Each median pools three rounds in each of five alternating process trials. Those fifteen observations are not fifteen independent processes.

Policy	Prefix bytes	Baseline	Integrated candidate	Change
byte fixed	0	100.906	83.641	-17.11%
byte variable	0	123.777	108.299	-12.50%
bit fixed	0	198.364	157.323	-20.69%
bit variable	0	239.904	200.643	-16.37%
byte fixed	64	106.890	85.060	-20.42%
byte variable	64	134.394	113.299	-15.70%
bit fixed	64	210.083	161.385	-23.18%
bit variable	64	263.059	206.662	-21.44%
byte fixed	4096	432.320	85.111	-80.31%
byte variable	4096	459.366	113.520	-75.29%
bit fixed	4096	529.348	166.295	-68.58%
bit variable	4096	567.505	209.307	-63.12%

The baseline and candidate ranges are disjoint in ten of twelve cases, including all four long-prefix fixtures. The overlapping cases are byte-fixed/prefix 0 and byte-variable/prefix 64. The 360 raw rows retain every round and its timing; metadata records full revisions, compiler commands and source/header hashes. A summary table gives minimum, median and maximum for every fixture in all five comparisons. The original key/value oracle and output checks passed for every row. These are measured resident-memory improvements on one host, not universal latency guarantees.

Framed-output comparison

The isolated comparison is 9c2c3be → 819578c. Its only production changes factor native frame output into one internal component and route the merger through it. The merger supplies the exact retained length and borrowed literal; the output keeps only the previous key's length. The two input cursors still retain decoded keys. The public native writer retains its own predecessor for order checking and shares the same rollback-safe framing primitive.

This removes the merger's third full-key allocation and changed-suffix copy. It does not eliminate output payload copies or the input decoders' key buffers. The following medians use the same twelve fixtures and trial schedule as above:

Policy	Prefix bytes	Previous writer	Framed output	Change
byte fixed	0	99.197	88.257	-11.03%
byte variable	0	119.914	113.742	-5.15%
bit fixed	0	176.524	161.494	-8.51%
bit variable	0	213.793	209.208	-2.14%
byte fixed	64	95.548	88.852	-7.01%
byte variable	64	120.750	112.361	-6.95%
bit fixed	64	177.055	160.080	-9.59%
bit variable	64	215.324	209.231	-2.83%
byte fixed	4096	96.036	86.899	-9.51%
byte variable	4096	123.197	113.673	-7.73%
bit fixed	4096	179.176	163.612	-8.69%
bit variable	4096	217.092	206.495	-4.88%

All twelve medians improve by 2.14–11.03%. Many minimum-to-maximum ranges overlap; the two 64-byte-prefix bit-fixed ranges are disjoint. These observations support the change on this host without establishing a universal speedup. The 360 raw observations preserve every range, and metadata pins the isolated comparison. All original-record and output-wire checks passed.

The complete 819578c checkpoint also passed strict Release/O3 ASan/UBSan native writer, allocation-failure rollback, native merge and mapped merge tests, plus installed and embedded package consumers (six checks). Those functional suites cover paused/moved builders, strict ordering, fixed and variable values, malformed later inputs, input pins and chronological composition. They are not timing data.

Frontier comparisons

The first run compares e45e47f with 13c914d, which carries each head's exact LCP with the previous output key and feeds a known retained length to the native writer. Its 4 KiB-prefix medians improve by 58.39–73.22%; the two zero-added-prefix byte fixtures regress by 8.82% and 10.45%. The complete first-run rows and metadata retain that observation.

The next run compares the same baseline with 32923e7. This additionally handles a successor's first differing policy unit directly, falling back to suffix comparison when a nonmaximal encoding repeats common content. Median nanoseconds per distinct output key are below. Each median pools three rounds in each of five alternating process trials; the fifteen observations are not fifteen independent processes. Negative change means less time.

Policy	Prefix bytes	Baseline	Frontier + first unit	Change
byte fixed	0	102.933	111.491	+8.31%
byte variable	0	133.631	135.099	+1.10%
bit fixed	0	210.617	193.059	-8.34%
bit variable	0	255.831	246.536	-3.63%
byte fixed	64	107.933	110.486	+2.37%
byte variable	64	133.815	137.009	+2.39%
bit fixed	64	214.457	194.005	-9.54%
bit variable	64	253.339	237.160	-6.39%
byte fixed	4096	447.871	112.292	-74.93%
byte variable	4096	467.087	135.885	-70.91%
bit fixed	4096	547.455	199.074	-63.64%
bit variable	4096	587.308	236.158	-59.79%

The raw rows include minimum-to-maximum spread through every recorded round; metadata records exact revisions, commands and execution order. Short fixtures have substantial overlapping ranges. The long-prefix reductions have disjoint baseline/candidate ranges in all four policies. The two frontier runs are separate paired experiments; their percentages are not multiplied together.

An additional 32923e7 → 819578c run is retained as raw observations and metadata. It crosses generic-EF and cold-exception integration as well as framing, so it is not the isolated framing comparison used above.

What the timer includes

Each fixture has 16,384 distinct output keys. A key consists of 0, 64 or 4,096 copies of byte p followed by an eight-byte big-endian integer. The bit policy adds three meaningful bits. Thus the zero-added-prefix case still has eight-byte keys, and adjacent integer encodings themselves share high zero bytes. Both policies use \(K=W=15\). Fixed values contain eight bytes or thirteen bits; variable values contain 8–24 policy units.

The two inputs each hold about two-thirds of the keys, with about one-third shared. Overlapping older values deliberately differ from the expected newer values, so replacement must choose the correct chronological source. Inputs and the batch oracle are constructed before timing. One warm merge is fully checked.

The measured interval constructs the merger and its two decoding cursors, runs step(128) until completion, builds the final Elias–Fano directory and destroys the merger. It excludes input construction, oracle validation and destruction of the returned output. The input owners, original fixture keys, expected output and warm output remain alive throughout. This is a resident-memory experiment; it does not establish cold-mapping, LLC-capacity, I/O or durable-publication cost. The timer divides by distinct output keys, not by total consumed input records.

Every measured result is checked after the timer. Its payload, metadata and every Elias–Fano word/sample are compared with batch encoding. Decoding is separately checked against the original integer-ordered full keys and their expected values. The batch encoder is shared library code; the original key/value oracle does not use merge comparisons to decide the result. The runner also requires identical payload sizes and complete-output fingerprints across variants and trials. This checks actual content, not only record counts or digest equality within one binary. The workload measures replacement; noncommutative composition, malformed sources, pause/move ownership and injected allocation failures have separate tests.

Reproduction and provenance

The host is an Apple M2 Max running macOS 26.6.2 and Apple Clang 21. Builds use C++20, -O3 -DNDEBUG -Wall -Wextra -Wpedantic -Werror. Each process requests QOS_CLASS_USER_INITIATED; this harness does not check that request's return code. CPU affinity and frequency are not fixed. Every compile and timing run uses the host's exclusive CPU-heavy and build-directory leases.

The shared runner materializes all headers from exact Git revisions and alternates baseline/candidate order across five process trials. The current harness and the first-run source snapshot differ only in untimed EF metadata access needed for the generic codec: the latter reads the old EF record_count, while the former reads the profile's record count and no longer compares that removed EF member. Their timed merge body is identical.

First-run harness SHA-256: def7ec472c20e2fb66530718196190d3011ed466413b943ccedcf72d07f21682.
Later harness SHA-256: d6d1825338c4c305dd269d704756d65a573533ae25cca2e13b28322d16019109.
Runner SHA-256: 44081c1fe74421b5c561ca8369ef8bcb6ea7889222bddbbd037c42c0e2049b81.

Run under the host resource gate when configured:

python3 bench/native_merge.py --baseline e45e47f --candidate 4429df0 \
  --build-dir build-native-merge-final --records 16384 --prefix 0 64 4096 \
  --rounds 3 --trials 5
python3 bench/native_merge.py --baseline 9c2c3be --candidate 819578c \
  --build-dir build-native-merge-framed --records 16384 --prefix 0 64 4096 \
  --rounds 3 --trials 5

The first comparison used the harness and runner at 13c914d, with candidate 13c914d. The metadata files retain every original input-header hash. The artifact check records executable hashes/sizes, verifies stored source/header hashes and checks matching payload sizes and output fingerprints across all 1,800 completed rows. No stage percentages are multiplied to infer a combined result: the headline comes from the direct final comparison above.

-Edward Kmett