Historical report imported 2026-10-10 from docs/experiments/benchmarks/native-resource-findings.md. Original SHA-256: 33de3bdd0dbc8d739188a4b67bd52c9e912fb889b56fc2766c56079f5d768c28. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Native frontier resource findings

The registered round (local checkout reference docs/experiments/benchmarks/native-frontier-protocol.md; not present at origin/main) completed its admitted native comparisons, diagnostic, verification and authenticated cleanup. Neither performance candidate is nominated. These records establish finite execution and preservation; they do not qualify service capacity or a deployment.

Accounting boundary

Totals below include final-closeout-01, and exclude the subsequent documentation recheck final-docs-02. That recheck has its own entry in the coordinator receipts (local evidence omitted from stage). Every job uses the resource launcher and serialized native coordinator on the mounted data drive. The enforced limits remain memory high/max 16/20 GiB, zero swap and an outer 1,800-second service deadline. Rust builds use two jobs; the last two full fast profiles also use two test threads. No remote workload ran.

Through closeout, 25 jobs charge 3,367.145781 seconds: 23 exited zero and two failed Clippy. Both failures remain charged. The new round retains 3,832.854219 seconds of its 7,200-second ceiling; the shared frontier retains 6,205.431596 seconds of 28,800. These are unused allowances, not authorization for unregistered work. No budget was reset. The prior campaign charges 16,595.017645 seconds, including its 14,383.687547-second frontier subset; the first continuation charges 946.699556 and the matched sweep 3,897.035519. The closeout's embedded ledger excludes its own 5.093782 seconds; the totals here add its completed receipt.

All recorded cgroup high/max/OOM counters and final swap use are zero. The largest job peak is 12,905,226,240 bytes from the initial full fast profile. The second and final full profiles peaked at 610,545,664 and 921,747,456 bytes with two test threads. These are whole-job observations with different cache/build contexts, not a matched benchmark of test concurrency or application RSS.

Job boundary	Elapsed seconds	Cgroup peak bytes	Outcome
Query screen	545.645	507,408,384	Exact chains; performance admission rejected
Ownership RSS diagnostic	278.928	615,301,120	Exact pairs; allocator-policy contrast supported
Candidate query regressions	200.892	1,838,968,832	75 passed, two ignored
Fast profile 1	384.565	12,905,226,240	Workspace tests passed; control-fixture Clippy failure
Fast profile 2	417.655	610,545,664	Workspace tests passed; example Clippy failure
Fast profile 3	426.013	921,747,456	All 17 checks passed

The complete commands, durations and resource records are retained in each coordinator job (local evidence omitted from stage), including CPU and IO counters. Build/test cgroup peaks, native RSS, requested live heap and timed query allocation are different measurement domains.

Verification and failures

Fast profile 3 (local evidence omitted from stage) and manual documentation profile 1 (local evidence omitted from stage) passed all seventeen and three checks respectively, with child/outer exit zero. The closeout summary (local evidence omitted from stage) binds all twenty receipts to the actual commands, revision and execution windows. Its positive control and fourteen rejected command/hash/path/window defects passed. The complete workspace test suite passed in both earlier failed fast profiles too; their failures were redundant outer OTLP request initializers. The separately registered control repair (local checkout reference docs/experiments/benchmarks/native-verification-retry-proposal.md; not present at origin/main) and example repair (local checkout reference docs/experiments/benchmarks/native-verification-example-retry-proposal.md; not present at origin/main) preserve those outcomes. A direct full-target Clippy preflight passed before profile 3.

Both failed outer trees were independently shown inactive, archived with every member, read back twice and re-inventoried before removal. Each contained 162 regular-file bytes plus directory structure; source and origin receipts accompany the archives. Eight controls, including unknown-job selection, were rejected in each preservation job. The first preservation (local evidence omitted from stage) and second preservation (local evidence omitted from stage) resolve cleanup without rewriting the historical failed scratch_removed flags. No failed original was deleted before its complete authenticated preservation.

Artifact disposition

Closeout authenticated six frozen binary copies twice before removing them: two memory executables on the data drive and four compressed query executables in the evidence directory, totaling 39,983,970 bytes. Their source snapshots, compiler settings and hashes remain. Successful native fixtures, decoder installs and temporary documentation runtimes were removed by their respective harnesses. The shared Cargo cache remains at $FABRIC_DATA/cargo. Bun hooks and CI remain disabled; documentation tools were invoked manually.

The completed closeout receipt charges memory evidence 7,794,688 bytes, query 98,799,616 and coordinator 18,976,768: 125,571,072 bytes total, below the 1 GiB aggregate and 512/384/128 MiB category caps. Its final coordinator receipt (local evidence omitted from stage) is the census boundary; later report-check artifacts are charged separately. Scratch admission retained the 8 GiB ceiling and 16 GiB free-space reserve. Sources, raw observations, complete query answer/control objects and failure archives remain; no historical evidence was discarded to admit this round.

Disposition

The ownership findings retain twelve exact paired results and failed RSS confirmation. The frozen-input diagnostic reproduced that excess under default allocation in both orders and reduced it with a fixed mmap threshold; it does not identify an allocation owner or nominate a production allocator setting. The query findings retain the primary benefit and 58 of 192 plain comparisons that fail at least one timing guard. Its 576 complete chains and 1,440 defective vectors were independently rechecked. Correctness success and performance admission remain separate.

The next research vectors require new registration and ledger admission. Consuming conversion, bounded buffer reuse, allocation-history modeling and state-sensitive query admission remain unimplemented hypotheses. No production default, wire format, durability order, new dependency or release qualification follows from this round.