The shared owning/mapped query path retains shape and bounds checks while internally built arrays avoid repeated padding and EF-endpoint reads. In the final comparison, median complete-query times are within 0.8% of the version before portable navigation views. This recovers the initial owning-path regression; it does not establish a new speedup over that earlier version.

What changed

The shared-view/navigation revision 3882d3c introduced byte-backed directory words, bounded rank accumulation, shared blob projection and generic query owners. Owning queries still constructed checked profile_view instances for their native and borrowed arrays at every catalog. Those constructors read the padding and first/terminal EF offsets each time.

The final candidate d8bc932 changes profile_array::view() to use profile_view::from_sections. Its private sections already come from checked builders. Each returned view still checks metadata and section shapes; it does not retain pointers across array copies or moves. The public profile_view constructor and explicit validate_contents() retain their content checks. Query parsing, rank overflow checks and projected-range bounds are unchanged.

The focused ownership tests cover valid default/built-empty arrays, nonempty moved-from shape rejection, independent copied allocations, moved destinations, replacement of the original owner and public malformed-EF rejection. Native and borrowed byte/bit roles are covered.

Method

Both experiments use the unchanged shared harness and runner, with the fixture and timer scope described in the complete-query comparison. Both variants use K=W=15. Each fresh process checks every expected source, native ordinal and full value before and after timing. The runner also checks catalog counts, match counts and value checksums across variants.

Each experiment has five alternating baseline/candidate process trials, three record/prefix cases, both byte and bit profiles, and 4,096 queries per case. The baseline in both experiments is 8214d4be528010aabd9aac76a0bcfb91ec7a5b14.

Initial candidate: 3882d3ccf91c5c3f964308b2693369917bae8034.
Final candidate: d8bc932600719e576cbdc9a8cffe4ead2da62dcc.

The final change is available on main at eafad7b, whose measured query and navigation headers match the final candidate. Reproduction uses that reachable revision.

Native M2 Max, AppleClang 21, C++20, -O3 -DNDEBUG, user-initiated QoS. All compilation and timing use the exclusive CPU/build-directory resource gate. There is no core pinning, fixed frequency or cold-storage measurement. Hardware probes denied by the local sandbox are recorded as unavailable in the metadata.

Initial regression

The first paired run found a 4.2–8.9% median regression. Times below are microseconds per complete query.

Records	Prefix bytes	Profile	Baseline	Shared views with repeated content checks	More query time
4,096	0	byte	3.129	3.409	8.94%
4,096	0	bit	6.131	6.457	5.30%
4,096	64	byte	3.129	3.403	8.73%
4,096	64	bit	6.208	6.471	4.23%
65,536	64	byte	3.782	4.110	8.66%
65,536	64	bit	7.632	7.996	4.77%

The initial CSV and metadata preserve all 60 rows, compiler commands, header/source hashes and process order.

Final owning-array path

The second experiment compares the final candidate directly with the same pre-view baseline. Its medians are effectively at parity. A negative change means less query time.

Records	Prefix bytes	Profile	Baseline µs/query	Final candidate	Change
4,096	0	byte	3.069	3.059	-0.32%
4,096	0	bit	6.056	6.012	-0.74%
4,096	64	byte	3.067	3.067	-0.02%
4,096	64	bit	6.070	6.035	-0.58%
65,536	64	byte	3.740	3.736	-0.12%
65,536	64	bit	7.593	7.596	+0.04%

Trial ranges overlap. The zero-prefix candidate includes a high trial, giving ranges of 3.029–3.612 µs for bytes and 5.965–6.617 µs for bits. These are ranges of trial averages, not per-query tail latencies. Preparation remains mixed and noisy because each process times one short root build; no general preparation improvement is claimed.

The final CSV and metadata contain every result. Comparisons above are within each paired run. The runs do not support a precise percentage attribution between the intermediate and final candidates.

All 120 release profile/trial rows passed the independent full-result and cross-variant checks. Encoded-array sizes are identical within each experiment. This accounting covers arrays reachable from one prepared root and excludes the separate warm root, capacities, fixed objects/view fields and allocation overhead; it does not establish unchanged total process memory. No encoded byte-hash equality check is claimed.

The final candidate also passed focused CMake ASan/UBSan profile and query suites plus the installed-package consumer check. The check record identifies exact source hashes and test configuration. These checks are independent of release timing.

Reproduction

Run each command under the host's CPU/build-directory resource gate. The runner snapshots headers from the supplied commits and compiles one shared harness.

python3 bench/query_compare.py --baseline 8214d4b --candidate 3882d3c \
  --build-dir build-shared-query-m2max --records 4096 --queries 4096 \
  --prefix 0 64 --larger-records 65536 --larger-prefix 64 --trials 5 --no-w16
 
python3 bench/query_compare.py --baseline 8214d4b --candidate eafad7b \
  --build-dir build-owning-view-m2max --records 4096 --queries 4096 \
  --prefix 0 64 --larger-records 65536 --larger-prefix 64 --trials 5 --no-w16

This benchmark measures owning in-memory queries. Mapped queries, page faults, disk I/O, recovery scans and native x86 performance are separate measurements.