The candidate typed-byte path has lower complete-path medians in 11/12 cases. Baseline/candidate ratios span 0.996–1.184×; 8 process-median ranges are disjoint in its favor. The raw-bit and typed-bit controls are mixed, so these results do not establish a general EF-cursor speedup.

The frozen comparison is 5868b32afae9e78cd86a71289ca119670d6c3748 → a760695c23c41deb4b8b64c2d7024e9919737b9a. The candidate combines forward EF navigation, byte-writer changes, and the built-in byte-string transport. The later equal-key donor improvement is outside this measurement, whose inputs have no equal keys or cancellations.

All modes represent the same sixteen-byte logical keys and present values. The typed paths are real, distinct transports: bit KV03 versus byte KV02. Raw KV02 controls remove typed framing and keep exactly the same raw key/value bytes. This is native merge throughput, not a complete transaction or indexed-world benchmark.

Method and spread

The collection contains 288 fresh processes, 864 timed iterations and 288 excluded warmups. Each case/mode/revision has three fresh processes with three measured iterations each. Tables show milliseconds: the median of those three process medians, followed by their minimum–maximum. Revision order alternates for each case/mode across rounds, with the unavoidable 2:1 first-slot split; mode and case order rotate. All trials are retained.

The complete timer includes merge setup on existing mapped inputs, native output allocation/encoding, EF construction, a fresh output mapping and direct section copy, body/header CRC32C, clipping and all return-time cleanup. Inputs are encoded, mapped and validated before timing. Fixture construction, sorted-record oracles and output validation are excluded. There is no fsync, msync, catalog publication or fractional-index construction. Inner phase diagnostics in the raw data omit some final destruction; the outer complete timer includes it.

Mode	Lower candidate medians	Baseline/candidate ratio range	Process ranges: faster / slower / overlap	All-trial ranges: faster / slower / overlap
typed-byte	11/12	0.996–1.184	8 / 0 / 4	5 / 0 / 7
raw-byte	11/12	0.996–1.180	8 / 0 / 4	3 / 0 / 9
raw-bit	6/12	0.977–1.092	1 / 1 / 10	0 / 0 / 12
typed-bit	7/12	0.969–1.068	0 / 1 / 11	0 / 0 / 12

typed-byte complete latency

Keys	Values	Records/input	Baseline ms	Candidate ms	Baseline/candidate	Process ranges
structured	FF16	4,096	0.731 [0.728–0.799]	0.620 [0.598–0.708]	1.178×	candidate-faster
structured	FF16	65,536	8.471 [8.360–8.599]	7.235 [7.158–7.361]	1.171×	candidate-faster
structured	FF16	262,144	33.348 [33.333–33.412]	28.154 [27.926–28.433]	1.184×	candidate-faster
structured	FV 0–512	4,096	1.283 [1.156–1.293]	1.199 [1.157–1.237]	1.070×	overlap
structured	FV 0–512	65,536	18.276 [18.199–18.498]	17.115 [16.731–18.941]	1.068×	overlap
structured	FV 0–512	131,072	38.946 [37.047–41.873]	34.646 [33.590–34.990]	1.124×	candidate-faster
hash-like	FF16	4,096	0.860 [0.815–0.877]	0.790 [0.780–0.833]	1.088×	overlap
hash-like	FF16	65,536	10.530 [10.456–10.660]	9.265 [9.172–9.270]	1.137×	candidate-faster
hash-like	FF16	262,144	42.165 [41.038–42.424]	37.028 [36.851–37.051]	1.139×	candidate-faster
hash-like	FV 0–512	4,096	1.387 [1.317–1.589]	1.393 [1.375–1.398]	0.996×	overlap
hash-like	FV 0–512	65,536	21.429 [21.104–21.958]	20.368 [20.151–20.934]	1.052×	candidate-faster
hash-like	FV 0–512	131,072	43.660 [42.626–43.718]	41.485 [40.951–42.144]	1.052×	candidate-faster

raw-byte complete latency

Keys	Values	Records/input	Baseline ms	Candidate ms	Baseline/candidate	Process ranges
structured	FF16	4,096	0.669 [0.661–0.711]	0.634 [0.607–0.636]	1.055×	candidate-faster
structured	FF16	65,536	7.314 [7.296–7.346]	6.601 [6.562–6.682]	1.108×	candidate-faster
structured	FF16	262,144	29.114 [29.045–29.579]	26.276 [26.070–26.614]	1.108×	candidate-faster
structured	FV 0–512	4,096	1.234 [1.200–1.257]	1.145 [1.109–1.178]	1.077×	candidate-faster
structured	FV 0–512	65,536	18.335 [17.670–18.732]	16.870 [16.397–16.895]	1.087×	candidate-faster
structured	FV 0–512	131,072	40.112 [36.627–41.203]	34.544 [34.136–37.314]	1.161×	overlap
hash-like	FF16	4,096	0.855 [0.830–0.922]	0.758 [0.729–0.849]	1.127×	overlap
hash-like	FF16	65,536	10.410 [10.387–10.603]	8.819 [8.660–8.830]	1.180×	candidate-faster
hash-like	FF16	262,144	41.202 [40.554–41.639]	35.475 [34.837–35.649]	1.161×	candidate-faster
hash-like	FV 0–512	4,096	1.447 [1.380–1.489]	1.275 [1.273–1.424]	1.135×	overlap
hash-like	FV 0–512	65,536	20.908 [20.552–21.081]	20.994 [20.086–21.205]	0.996×	overlap
hash-like	FV 0–512	131,072	44.346 [42.732–48.234]	41.372 [41.292–42.346]	1.072×	candidate-faster

raw-bit complete latency

Keys	Values	Records/input	Baseline ms	Candidate ms	Baseline/candidate	Process ranges
structured	FF16	4,096	0.999 [0.997–1.077]	1.009 [0.972–1.026]	0.990×	overlap
structured	FF16	65,536	12.100 [11.845–12.217]	12.389 [12.289–12.463]	0.977×	candidate-slower
structured	FF16	262,144	48.775 [47.856–49.084]	48.933 [48.905–49.385]	0.997×	overlap
structured	FV 0–512	4,096	2.099 [1.817–2.178]	1.923 [1.854–1.955]	1.092×	overlap
structured	FV 0–512	65,536	28.923 [28.540–31.710]	28.952 [28.614–30.097]	0.999×	overlap
structured	FV 0–512	131,072	59.407 [58.502–66.444]	58.586 [58.174–65.307]	1.014×	overlap
hash-like	FF16	4,096	1.347 [1.305–1.410]	1.358 [1.318–1.400]	0.992×	overlap
hash-like	FF16	65,536	18.545 [18.206–18.660]	18.307 [18.267–18.414]	1.013×	overlap
hash-like	FF16	262,144	73.568 [73.154–74.881]	73.700 [72.123–74.440]	0.998×	overlap
hash-like	FV 0–512	4,096	2.197 [2.182–2.240]	2.175 [2.158–2.182]	1.010×	candidate-faster
hash-like	FV 0–512	65,536	33.939 [33.585–34.993]	33.338 [32.947–33.762]	1.018×	overlap
hash-like	FV 0–512	131,072	68.475 [68.144–79.366]	67.982 [67.945–68.511]	1.007×	overlap

typed-bit complete latency

Keys	Values	Records/input	Baseline ms	Candidate ms	Baseline/candidate	Process ranges
structured	FF16	4,096	2.052 [1.959–2.140]	2.019 [2.011–2.043]	1.016×	overlap
structured	FF16	65,536	30.238 [29.250–30.255]	29.844 [29.650–30.682]	1.013×	overlap
structured	FF16	262,144	121.347 [121.006–121.421]	122.144 [121.723–122.500]	0.993×	candidate-slower
structured	FV 0–512	4,096	3.163 [2.898–3.280]	2.962 [2.905–3.083]	1.068×	overlap
structured	FV 0–512	65,536	47.301 [47.166–47.334]	48.116 [45.911–49.447]	0.983×	overlap
structured	FV 0–512	131,072	96.031 [95.716–99.223]	94.728 [94.510–100.304]	1.014×	overlap
hash-like	FF16	4,096	2.277 [2.222–2.332]	2.304 [2.280–2.330]	0.988×	overlap
hash-like	FF16	65,536	34.651 [34.419–35.218]	34.231 [33.783–35.390]	1.012×	overlap
hash-like	FF16	262,144	143.341 [140.661–146.616]	140.463 [139.848–142.466]	1.020×	overlap
hash-like	FV 0–512	4,096	3.224 [3.163–3.317]	3.327 [3.199–3.399]	0.969×	overlap
hash-like	FV 0–512	65,536	52.063 [50.956–52.228]	53.107 [51.714–53.255]	0.980×	overlap
hash-like	FV 0–512	131,072	105.415 [104.689–105.946]	105.089 [104.558–106.312]	1.003×	overlap

Control interpretation

The raw-bit structured FF16 case at 65,536/input has a 2.39% higher candidate median with disjoint process-median ranges. Its all-trial range classification is overlap.

The typed-bit structured FF16 case at 262,144/input has a 0.66% higher candidate median with disjoint process-median ranges. Its all-trial range classification is overlap.

The raw-bit results are a useful control for the generic profile navigation change. The unchanged typed-bit path also varies. The small regressions have not been causally isolated. Source inspection shows the forward EF cursor copies its complete view/state only at codec boundaries (once per fifteen records here), then commits that copy after parsing succeeds; it does not copy the whole cursor on every record. That is a possible follow-up cost to investigate, not a measured explanation. No additional experiment or selective rerun was used to remove these results.

Candidate bit versus byte

These cross-mode ratios use the same frozen candidate and logical records. They compare the actual complete formats, including their different framing and output sizes. They do not isolate one instruction or one wire-code feature.

Keys	Values	Records/input	Typed bit ms	Typed byte ms	Typed bit / byte	Raw bit / byte
structured	FF16	4,096	2.019	0.620	3.255×	1.591×
structured	FF16	65,536	29.844	7.235	4.125×	1.877×
structured	FF16	262,144	122.144	28.154	4.338×	1.862×
structured	FV 0–512	4,096	2.962	1.199	2.471×	1.679×
structured	FV 0–512	65,536	48.116	17.115	2.811×	1.716×
structured	FV 0–512	131,072	94.728	34.646	2.734×	1.696×
hash-like	FF16	4,096	2.304	0.790	2.915×	1.792×
hash-like	FF16	65,536	34.231	9.265	3.695×	2.076×
hash-like	FF16	262,144	140.463	37.028	3.793×	2.078×
hash-like	FV 0–512	4,096	3.327	1.393	2.389×	1.705×
hash-like	FV 0–512	65,536	53.107	20.368	2.607×	1.588×
hash-like	FV 0–512	131,072	105.089	41.485	2.533×	1.643×

Exact encoded output sizes

Sizes are complete logical file bytes, including envelope, directory, EF, padding and checksums. Raw-bit, raw-byte and typed-bit files are byte-identical between revisions; only typed-byte wire changes. All input file sizes and hashes are also retained in the raw checks.

Keys	Values	Records/input	Raw bit	Raw byte	Typed bit	Typed byte baseline	Typed byte candidate
structured	FF16	4,096	142,528	156,960	152,888	198,656	173,544
structured	FF16	65,536	2,276,680	2,507,608	2,441,152	3,174,520	2,772,816
structured	FF16	262,144	9,105,952	10,029,680	9,763,608	12,697,296	11,090,488
structured	FV 0–512	4,096	2,130,192	2,137,376	2,125,240	2,174,584	2,145,584
structured	FV 0–512	65,536	34,086,264	34,201,176	34,005,736	34,796,392	34,332,512
structured	FV 0–512	131,072	68,171,576	68,401,400	68,010,440	69,591,824	68,664,072
hash-like	FF16	4,096	277,376	271,048	287,712	312,592	287,512
hash-like	FF16	65,536	4,370,152	4,262,496	4,534,352	4,926,640	4,525,832
hash-like	FF16	262,144	17,350,048	16,880,552	18,005,776	19,536,120	17,933,976
hash-like	FV 0–512	4,096	2,264,776	2,251,296	2,259,808	2,288,464	2,259,504
hash-like	FV 0–512	65,536	36,175,696	35,953,440	36,094,904	36,547,840	36,084,776
hash-like	FV 0–512	131,072	72,318,160	71,855,496	72,156,216	73,044,048	72,118,160

The new typed-byte transport reduces complete file size by 1.27–12.65% relative to the old typed-byte transport across these fixtures. Against typed-bit, the space tradeoff depends on the key distribution and value sizes:

structured FF16: candidate typed-byte uses 13.511–13.590% more complete file space than typed-bit.
structured FV 0–512: candidate typed-byte uses 0.957–0.961% more complete file space than typed-bit.
hash-like FF16: candidate typed-byte uses 0.070–0.399% less complete file space than typed-bit.
hash-like FV 0–512: candidate typed-byte uses 0.013–0.053% less complete file space than typed-bit.

The following ratios use complete output bytes for the same records; a ratio below one means the numerator occupies less space. Total file bits are exactly eight times the byte counts above. The full space table also records those bit totals, logical key/value byte counts, and component sizes for all sixty distinct case/format/revision combinations.

Keys	Values	Records/input	Candidate typed byte / typed bit bytes	Candidate / baseline typed byte bytes
structured	FF16	4,096	1.135105	0.873591
structured	FF16	65,536	1.135864	0.873460
structured	FF16	262,144	1.135901	0.873453
structured	FV 0–512	4,096	1.009573	0.986664
structured	FV 0–512	65,536	1.009609	0.986669
structured	FV 0–512	131,072	1.009611	0.986669
hash-like	FF16	4,096	0.999305	0.919768
hash-like	FF16	65,536	0.998121	0.918645
hash-like	FF16	262,144	0.996012	0.917991
hash-like	FV 0–512	4,096	0.999865	0.987345
hash-like	FV 0–512	65,536	0.999719	0.987330
hash-like	FV 0–512	131,072	0.999473	0.987324

Space components

These are exact physical section lengths from the retained file headers, with no new merge runs. The front-coded record stream includes key controls, key suffixes and values; it is not solely user payload. EF includes low/high words, select samples, sparse entries and their word padding. Other bytes comprise the envelope, directory, sort metadata where present, and section alignment. The three physical columns sum to the complete file size. The separate FC bit count excludes its final partial-byte padding.

The largest case for each key/value distribution is shown here: case 2 is structured FF16, case 5 structured FV, case 8 hash-like FF16, and case 11 hash-like FV. FF cases use 262,144/input and FV cases 131,072/input; the linked table contains every size.

Case	Format	Revision	FC logical bits	FC storage bytes	EF bytes	Other bytes	Total bytes
2	raw-bit	candidate	72,506,359	9,063,295	42,432	225	9,105,952
2	raw-byte	candidate	79,958,104	9,994,763	34,688	229	10,029,680
2	typed-bit	candidate	77,766,721	9,720,841	42,448	319	9,763,608
2	typed-byte	baseline	101,194,928	12,649,366	47,704	226	12,697,296
2	typed-byte	candidate	88,346,712	11,043,339	46,920	229	11,090,488
5	raw-bit	candidate	545,066,468	68,133,309	38,040	227	68,171,576
5	raw-byte	candidate	546,957,360	68,369,670	31,504	226	68,401,400
5	typed-bit	candidate	543,776,762	67,972,096	38,032	312	68,010,440
5	typed-byte	baseline	556,480,176	69,560,022	31,576	226	69,591,824
5	typed-byte	candidate	549,058,592	68,632,324	31,520	228	68,664,072
8	raw-bit	candidate	138,326,960	17,290,870	58,952	226	17,350,048
8	raw-byte	candidate	134,679,352	16,834,919	45,408	225	16,880,552
8	typed-bit	candidate	143,572,090	17,946,512	58,952	312	18,005,776
8	typed-byte	baseline	155,881,952	19,485,244	50,648	228	19,536,120
8	typed-byte	candidate	143,067,960	17,883,495	50,256	225	17,933,976
11	raw-bit	candidate	578,237,302	72,279,663	38,272	225	72,318,160
11	raw-byte	candidate	574,588,472	71,823,559	31,712	225	71,855,496
11	typed-bit	candidate	576,941,042	72,117,631	38,272	313	72,156,216
11	typed-byte	baseline	584,096,608	73,012,076	31,744	228	73,044,048
11	typed-byte	candidate	576,689,704	72,086,213	31,720	227	72,118,160

Validation and provenance

All 288 processes passed exact whole-file equality with an independently ordered, full-record writer oracle after every iteration. Every output received a full CRC/semantic scan; warmup outputs and all inputs were independently decoded to the logical records.
All canonical logical input/output SHA-256 values match across four modes, both revisions, all processes, and the earlier matched Metal fixtures. Only the typed-byte wire change is allowed across revisions.
Independently checked 864 retained physical headers for native layout kind, record count, physical extent and header CRC32C. Full physical SHA-256 values are stable within each format/revision. Disposable full input/output files were removed after checks and header/hash capture.
Both release builds passed sixteen serial CTests. The candidate also passed all sixteen strict O2 ASan/UBSan CTests. Only the O3 release binaries supplied the retained timings.
The collector verified both immutable header revisions (82 files each), 5 harness/fixture sources, and both binary hashes before and after collection. Measured harness source: 40aecd43f21c8c4a16ebd3978a19042b61779d5f.
One preflight attempt stopped while reading sandbox-restricted CPU identity metadata, before any benchmark process. The identical frozen collector then ran with authorized read-only metadata access; no timed attempt was replaced.

Host: Apple M2 Max, macOS 26.6.2 (25G83). Collection date: 2026-09-16. C++20 with O3, NDEBUG and strict warnings. The existing fixture limit of 131,072/input for FV keeps the grid identical to the Metal comparison; it is not a CPU format-size limit.

See raw rows, all summaries, source/artifact manifest, logical/physical checks, schedule/host metadata, correctness qualification, and the method. Earlier Metal calibration and fixed-format measurements are unchanged.