The final implementation removes the initial byte-query regressions while retaining the smaller version-2 encoding and avoiding predecessor-block reads for routed comparison repair. Against the original implementation, complete-query medians range from 3.04% faster to 0.91% slower. Most trial ranges overlap. Preparation is mixed, including a 20.97% slower small byte-policy median; its trial ranges overlap too.

Final complete-query comparison

Original: 18daa213552ec44872f9c042a7977fbe7c81c07d.
Final measured candidate: 7fe6ee841a496497057d31a7c1e291c9d1d8f02c.
Reachable integration: 030a7a12bbfeaf395d7f1ae9f27eaf4729372b4c.

Every Everett header in the recursive query.h include closure is byte-identical between the measured and reachable candidates. The reproduction record lists those hashes; unrelated merge, file-writing and catalog code is outside the harness's include closure.

Median nanoseconds per complete query, five trials; lower is better.

Base records	Prefix bytes	Policy	Original	Final	Change
4,096	0	byte	1,852.987	1,869.883	+0.91%
4,096	0	bit	2,819.773	2,734.080	−3.04%
4,096	64	byte	1,827.026	1,814.819	−0.67%
4,096	64	bit	2,869.151	2,815.633	−1.87%
65,536	64	byte	2,248.464	2,222.890	−1.14%
65,536	64	bit	3,477.315	3,438.558	−1.11%

Only the small 64-byte-prefix bit case has disjoint query ranges. All 60 final rows pass the full-result oracle and cross-revision semantic checks. No case retains the initial regression above 5% with disjoint ranges. This is one direct comparison; percentages from the intermediate experiments are not added to it.

Median microseconds to prepare the root from an already-built chain:

Base records	Prefix bytes	Policy	Original	Final	Change
4,096	0	byte	7.458	8.458	+13.41%
4,096	0	bit	9.792	10.000	+2.12%
4,096	64	byte	7.750	9.375	+20.97%
4,096	64	bit	10.917	10.958	+0.38%
65,536	64	byte	65.875	64.584	−1.96%
65,536	64	bit	110.292	106.750	−3.21%

Every preparation range overlaps. These short single-build measurements show the median differences but do not isolate their cause or establish a general preparation speedup.

What changed

The first frame of a physical FC block now stores its retained-prefix position directly. Later frames still backspace from their physical predecessor. This removes the block's separate predecessor-length checkpoint and lets an encoded cursor traverse literals and values without reconstructing keys. The profile encoding version is 2.

At a routed cut, the exact cut LCP repairs the outgoing comparison without reading the preceding borrowed block. A comparison context can therefore have an unknown full key length; full_units() reports that honestly with an optional. The optional state fits existing padding: both policy contexts remain 40 bytes on this host. Encoded records shrink from 112 to 96 bytes, and profile views remain 272 bytes. The size probe preserves its source, commands and exact results.

I checked the framing change with the existing query harness and runner, whose bytes are unchanged between the original baseline and the first version-2 candidate. Queries still compare full results against the same independent integer-key oracle. Wire bytes differ, so semantic agreement and backing-array sizes are checked separately.

Navigation work

The untimed audit runs the same fixture and checks every source, ordinal and value. For \(K=W=15\), the original and initial version-2 paths have identical record visits, skipped headers and compared-bit counts. Totals over 4,096 queries are:

Base records	Native skipped / visited	Borrowed skipped / visited
4,096	89,389 / 120,034	73,481 / 52,020
65,536	85,071 / 120,158	85,704 / 88,257

These counts are the same for both policies and the measured prefixes. They count record headers, not individual count-code decodes or CPU instructions. The new route carries an unknown length 1,070 times in each small case and 1,305 times in each large case. This fixture performs no extra false-borrow value probes; independent protected-page tests exercise that recovery path and preceding-block avoidance. Raw work rows and metadata retain the separate bit comparison counts and the \(W=16\) control. The final counter check repeats all six \(W=15\) cases on the final parser and requires every reported field to equal the initial version-2 audit; it passes. Its metadata pins the final headers and exact source.

Initial regression and isolated changes

Before measuring, I used a practical investigation threshold: a median regression above 5% with non-overlapping trial minimum/maximum ranges. This is a debugging trigger, not a statistical confidence interval. The first version-2 candidate crossed it for both byte-policy cases with a 64-byte prefix:

Base records	Prefix bytes	Policy	Original ns/query	Initial version 2	Change
4,096	0	byte	1,917.328	2,063.182	+7.61%
4,096	0	bit	2,917.236	3,059.163	+4.87%
4,096	64	byte	1,927.012	2,103.902	+9.18%
4,096	64	bit	3,057.526	3,183.054	+4.11%
65,536	64	byte	2,290.883	2,516.408	+9.84%
65,536	64	bit	3,591.268	3,730.591	+3.88%

I preserved each subsequent comparison rather than adding its percentage to another experiment:

Returning scalar location state instead of carrying a complete encoded frame through the pre-lane loop reduced medians by 2.03–3.68% in its paired run.
Bounding the absolute retained prefix once against the physical frame offset permits ordinary key-length addition after payload bounds checks. Its isolated timing was mixed, from −3.87% to +4.89%, with overlapping ranges throughout.
A direct comparison including those two changes removed the larger byte regression, but the small 64-byte-prefix case still regressed 5.59% with disjoint ranges.
The next change gives skipped frames a scalar parsing path. It shares the count and payload checks with full parsing, but returns only key length and next offset. It does not construct suffix/value views or a complete encoded record for data that the search will skip. Its paired comparison improves all six medians by 1.35–4.72%; the small 64-byte-prefix byte and bit ranges are disjoint. The final table above compares the complete result directly with the original implementation.

The absolute-prefix bound uses physical offsets after fixed-value stride restoration. Every inherited unit appeared in an earlier literal, so the retained prefix cannot exceed that position. Relative frames inherit the same bound from the previous frame's end. Once suffix and value lengths fit the admitted extent, the key length also fits it. Tests include impossible absolute prefixes, maximal count operands and malicious EF offsets into literal data.

Array sizes

For \(K=W=15\), the format change saves approximately 0.4–0.6% of the encoded arrays in this fixture. Parsing changes preserve these sizes.

Base records	Prefix bytes	Policy	Original bytes	Version-2 bytes
4,096	0	byte	62,355	61,961
4,096	0	bit	59,310	58,943
4,096	64	byte	62,867	62,473
4,096	64	bit	60,118	59,847
65,536	64	byte	1,101,686	1,095,462
65,536	64	bit	1,071,470	1,067,110

These are encoded arrays reachable from one prepared root, counting shared native arrays once. They include native and borrowed FC payload, both EF directories, rank classes/checkpoints, cut-LCP arrays and false-borrow flags. They exclude capacities, object headers, allocators, fixture/oracle keys, query scratch and the separately retained warm root. They are not total process memory or storage-device traffic.

The initial version-2 run also compared \(W=16\) with \(K=15\). It saved a small additional number of directory/payload bytes but had slower query medians in all six cases. This control uses the initial parser, before the scalar fixes; it does not establish the final parser's \(W=16\) performance.

Base records	Prefix bytes	Policy	\(W=15\) ns/query	\(W=16\) ns/query	\(W=16\) array bytes
4,096	0	byte	2,063.182	2,223.175	61,937
4,096	0	bit	3,059.163	3,242.645	58,837
4,096	64	byte	2,103.902	2,201.101	62,449
4,096	64	bit	3,183.054	3,362.915	59,714
65,536	64	byte	2,516.408	2,837.168	1,095,046
65,536	64	bit	3,730.591	4,310.425	1,064,879

Method and evidence

The host is an Apple M2 Max with Apple Clang 21. Release runs use C++20, -O3 -DNDEBUG and strict warnings. Five fresh-process trials alternate revision order. Each process warms the exact-result oracle, measures preparation and 4,096 complete queries, then checks all results again. Preparation includes root validation and any empty-native routing prefix; it excludes construction of the four native levels and their initial indexes. The fixture has \(N\), \(N/4\), \(N/16\) and \(N/64\) native records, fixed prefix bytes and ordered integer suffixes; the bit policy adds a parity bit.

Every returned value contributes to the timed checksum. Revision comparisons also require identical match counts, catalog visits and prepared shape. The harness verifies its QoS request. CPU/build-directory leases exclude other participating heavy jobs; core assignment and frequency are not fixed. These are resident owning queries, without cold mmap faults, storage or network I/O.

All raw observations, including scheduling outliers, remain in the linked files. Metadata records exact revision pins, every header hash, the shared source and runner hashes, compiler commands and process order:

Comparison	Raw rows	Metadata
Original `18daa21` → initial version 2 `e042920`, including \(W=16\)	90 rows	Metadata
Initial `e042920` → scalar location `17dff1c`	60 rows	Metadata
Scalar location → absolute bound `a12eb80`	60 rows	Metadata
Original → location plus bound	60 rows	Metadata
Bound → scalar skipped frames `7fe6ee8`	60 rows	Metadata
Original → final combined candidate	60 rows	Metadata
Original → initial version 2, ASan/UBSan	12 rows	Metadata

The artifact manifest records the sizes and SHA-256 hashes of the source, rows and metadata files.

Sanitizer timings are not performance evidence. The final focused profile, comparison and complete-query suites pass 3/3 with -O3 and ASan/UBSan. They cover byte/bit keys, fixed and variable values, independent \(K/W\) combinations, ordinary and explicit-restart streams, partial comparison contexts, malformed framing and protected payload pages. The check record retains flags, exact source hashes and the complete test output.

To repeat the final comparison from reachable revisions:

python3 bench/query_compare.py --baseline 18daa21 --candidate 030a7a1 \
  --build-dir build-query-absolute --records 4096 --queries 4096 \
  --prefix 0 64 --larger-records 65536 --larger-prefix 64 --trials 5 --no-w16

Run under the host resource lease when configured. The runner removes old snapshot directories before materializing pinned headers. Sanitizer smoke tests use --sanitize --records 64 --queries 128 --larger-records 0 --trials 1; their elapsed times are not query-performance measurements.