Historical report imported 2026-10-10 from docs/experiments/benchmarks/catalog-boundary-findings.md. Original SHA-256: 9c61c1e5210ef30867b70d3f35d8fe12fb41b1da612442f6f53d0aab892babaa. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Storage read boundary comparison

The finite CQ1 comparison completed: 18 trials, 1,152 exact full-chain oracle acceptances and all three registered plain performance guards met. This supports the narrow internal ownership boundary plus discovery correction on the registered workload. It does not establish service throughput or universal query overhead. Protocol (local checkout reference docs/experiments/benchmarks/catalog-boundary-protocol.md; not present at origin/main), archive amendment (local checkout reference docs/experiments/benchmarks/catalog-boundary-archive-protocol.md; not present at origin/main), consolidation (local evidence omitted from stage).

The candidate moves Walk's physical source/index/cache ownership into read_catalog.rs; Scan physical discovery remains in History. Query predicates, sort keys, pagination and domain semantics remain in query/core. The comparison also includes the publication/reclaim discovery consistency correction, so it cannot isolate module movement alone. Default descriptor ownership stays cloned; shared metadata is an explicit experimental selector.

Plain paired evidence

Ratios below are candidate / baseline, using each trial's median of three warm broad-tail Walk first-page measurements. The native pre-continuation VmHWM guard includes fixture and retained answer allocations; it is not catalog-only heap. The guard is at most 1.05 for all three metrics in every pair.

Pair	Corrected catalog wall	Corrected catalog CPU	Native HWM	Guard
1	0.95897	0.96021	0.99642	Met
2	0.98636	0.98501	0.99603	Met
3	1.01854	1.02066	0.99303	Met

The direction varies; the result is preservation within the declared guard, not a claimed general speedup. Shared/clone one-Segment wall ratios were 0.98377, 1.00951 and 0.99064; CPU ratios were 0.98233, 1.01072 and 0.99065. Those small mixed differences do not inherit the isolated metadata probe's large speedup. CR3's many-Segment/concurrent-reader comparison remains separate.

Six 128-row preflights cover plain/counted and baseline/clone/shared. Full trials use 65,536 seed-42 shuffled logs, 1,024-byte bodies, four query shapes, tail and one-Segment layouts, Scan/Walk, first plus three warm calls and limit 10,000. Every timed first page is associated with its actual complete continuation chain. Source/timestamp-rank/Batch identities match across slices; raw timings, phase archives, complete answer maps, ledger objects and mutation controls are retained. Counted results are diagnostics, never substituted for uninstrumented timings.

Frozen baseline hashes are in the original protocol. Candidate plain SHA256 is 417beb5e07623f8752b3bb0e93017f514c7f0ecc29139150d73a3f9951106dad; counted is 83c3381cdd100155db9dd71be1dac089b04d7f2169c259fb72ae40d8ba4ddf39. All later slices restored these exact executables. Formatting/test-fixture/lint cleanup after capture does not retroactively change their source attribution.

Commands, resources and archive limits

All commands are the protocol's resource-launcher/coordinator commands, with literal arguments retained in each receipt. Diagnostic and pairs 1/2/3 exited 0 in 411.686 / 366.057 / 361.024 / 363.070 seconds. Whole-job peaks rounded by systemd were 4.2 / 4.4 / 4.3 / 4.2 GiB, including grading; swap stayed zero. Successful owned scratch was removed. catalog-boundary-summary-01 independently recomputed raw medians/HWM/ratios, checked complete unique chain identities and matched frozen provenance; exit 0 in 1.003 seconds.

The initial per-slice archive accounting exceeded the planned aggregate cap. Exact compaction preserved every decoded byte but still exceeded 256 MiB, so catalog-evidence-compaction-01 remains exit 2. The separately registered allocation review (local checkout reference docs/experiments/benchmarks/catalog-evidence-allocation-protocol.md; not present at origin/main) permits 1 GiB query evidence; compaction run 02 passed. Later slices share exact objects and compress large closed JSONs with verified readback. Logical path bytes and unique stored bytes remain distinct. Memory/scratch/swap/time bounds were never increased.

Separate discovery, Scan, paused cuts and cold/lifetime controls cover selected real transitions. They do not prove all concurrent schedules. Cached manifests assume immutable publication and no label reuse; metadata ownership never leases files or extends expired snapshot rights. Native ingress/Spool acceptance, H256, large-row output pressure and deployment qualification are outside CQ1.