I compared 9c2c3be with 4429df0 using the unchanged complete-query harness and runner. The result is broadly flat: these six fixtures show no clear material query regression from the integrated changes. I do not infer a general speedup from spelling arithmetic as shifts and masks.

The candidate includes explicit bit/byte shifts and masks, add-before-rounding with admission bounds, packed rank spacer changes, and shared native frame output. This is an integration comparison; it does not isolate the cost of any one check or source-level expression. In particular, the new bounds in bit_string::validate and profile_detail::resize are present in the measured candidate.

Method

Both revisions compile one byte-identical harness with AppleClang 21, C++20, -O3 -DNDEBUG for native arm64 on the local M2 Max host. The metadata records the compiler and platform; sandbox restrictions prevented the optional CPU-brand, memory-size and CPU-count probes. Both builds and all timed processes held the host's exclusive CPU/build-directory lease.

I used K=W=15, variable values, 4,096 queries per trial, and five fresh-process trials for each case. Baseline and candidate execution order alternates; each process measures the byte profile before the bit profile. Four native catalogs contain N, N/4, N/16 and N/64 entries. The prepared root adds one routing catalog for N=4096 and two for N=65536.

Keys contain an eight-byte big-endian integer after zero or 64 shared prefix bytes; bit-profile keys add a meaningful parity bit. The independent integer oracle checks exact source identity, ordinal and complete value before and after timing. Timed work includes cursor construction/destruction, query and result ownership, catalog navigation, and a checksum consuming every returned value byte. It excludes original chain construction. Root preparation has its own separate timer.

Query results

Values are microseconds per complete query. The change column compares medians; negative means less elapsed time. Ranges show the minimum and maximum of five trials, not confidence intervals.

Base entries	Prefix bytes	Profile	Baseline median	Candidate median	Change	Baseline range	Candidate range
4,096	0	byte	2.706	2.654	−1.9%	2.692–2.745	2.635–2.677
4,096	0	bit	4.091	4.123	+0.8%	4.068–4.154	4.070–4.180
4,096	64	byte	2.670	2.669	−0.0%	2.640–2.704	2.624–2.688
4,096	64	bit	4.184	4.173	−0.3%	4.149–4.193	4.155–4.198
65,536	64	byte	3.406	3.225	−5.3%	3.317–3.486	3.161–3.477
65,536	64	bit	5.155	5.264	+2.1%	5.085–5.334	5.078–5.546

The largest median increase is 2.1% in the larger bit case, whose ranges overlap broadly. The larger byte median falls by 5.3%, but its ranges also overlap. The small zero-prefix byte case has separate ranges and a modest 1.9% decrease. I would keep the cleanup on this evidence without attributing these differences to individual instructions. Five warm trials do not establish cold-mapping, x86, disk-I/O or general workload performance.

Preparation is mixed and measures just one short root build per process:

Base entries	Prefix bytes	Profile	Baseline preparation µs	Candidate preparation µs
4,096	0	byte	17.50	18.83
4,096	0	bit	27.50	27.04
4,096	64	byte	25.71	16.46
4,096	64	bit	23.83	25.83
65,536	64	byte	112.92	105.04
65,536	64	bit	171.33	158.67

I make no general preparation-speed claim from these short measurements.

Correctness, storage and reproduction

All 60 profile/trial rows passed their integer oracle and cross-revision checks for head sizes, routing depths, catalog visits, match counts and checksums. Each small-case trial returned 2,796 matches over 20,480 catalog visits; the larger case returned 2,810 matches over 24,576 visits. Every counted payload and auxiliary-array component is identical between the two revisions. The harness does not compare the complete encoded bytes.

The raw CSV retains every trial and storage component. The metadata retains resolved revisions, all header hashes, unchanged harness/runner hashes, compiler commands, process order and environment probes. Counted storage excludes capacities, fixed objects, allocation overhead, retained warm roots and query scratch.

Run the unchanged runner under the host's exclusive CPU/build-directory gate:

python3 bench/query_compare.py \
  --baseline 9c2c3beeb0f5ce6cf3d81eb5963d41efdc22041e \
  --candidate 4429df0a8507705a9c4ccc0f6366344cf543f396 \
  --no-w16 --records 4096 --queries 4096 --prefix 0 64 \
  --larger-records 65536 --larger-prefix 64 --trials 5 \
  --build-dir build-query-rounding

The recorded harness SHA256 is 5200a36f81c194c854443a3e031db7eb6158062df88f77829fb9e8049269df18; the runner SHA256 is 92c73c1a5eeb01d50919178ff272541283beb5445730d2b50b18af63cde21fd3.