Historical report imported 2026-10-10 from docs/experiments/benchmarks/coupled-labs-run-01.md. Original SHA-256: c3ae40c98fd4a5531ab3ab1814fe3724b681b92f212ed35379df373a1bc1a450. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Coupled labs, run 01

Status: admitted experiments complete; correctness gates remain open, 2026-10-07. Owner authorized launch. The plan (local checkout reference docs/experiments/benchmarks/coupled-labs-plan.md; not present at origin/main), queue (local checkout reference docs/experiments/benchmarks/coupled-lab-queue.json; not present at origin/main) and separately registered initial admission (local checkout reference docs/experiments/benchmarks/coupled-launch-protocol.md; not present at origin/main) apply. Root admits one contained workload at a time; lab PIs prepare source/fixtures independently.

Operations

O6 catalog-coupled-o6-reproduce-01 ran the unchanged filtered completion_storage missing-query-table test, all features, offline/locked, through the launcher, coordinator and resource admission helper. Exit 101, 4.006 seconds; rounded whole-job peak 104 MiB, zero swap. Zero expected/answered rows agree, but retained window, freshness and gaps fail the unchanged oracle. This establishes the current counterexample, not the correct resolution of its availability-model disagreement. No oracle, contract or expected outcome changed. Source archive, stderr and admission observations remain in the coordinator job.

catalog-coupled-o6-cleanup-01 exited 0 in 0.503 seconds, 33.9 MiB rounded peak, zero swap, owned scratch removed. The existing cleanup helper archived and byte-verified the failed fixture before removing it. Original failure stays failed.

O7 catalog-coupled-o7-ledger-01 exited 0 in 3.505 seconds. Ordered raw/log/metric hashes and logical counts agreed for the historical 4070-log/6400-metric fixture. Its different physical layouts admitted hypothetical footer counts of 20350 and 9156 rows for the overlapping window, while both matched 7140 rows. These are not measured I/O counts; absent per-group row identities/bounds prevent a soundness conclusion. Negative controls were rejected. The original C2 gate remains failed. C5/O8 remain behind their declared correctness and resource dependencies. No experimental mechanism is promoted.

Resources

Registration commit 26a5e9c allocates a prospective 832 MiB capacity allowance for subsequent coupled jobs within unchanged 2 GiB aggregate; the earlier 768 MiB violation remains recorded. The 192 MiB incremental envelope includes source, receipt and failure overhead. Initial admission observed 1,696,272,384 allocated aggregate bytes and 815,308,800 mandatory capacity bytes. New workloads remain 20 GiB max/no swap, data-drive scratch and existing frontier remainder. The resource helper has injected length/block overage controls, pre/post inventory, and propagates the original child exit. Future native protocols are registered only after their real commands and controls exist.

Capacity C4

Findings and result (local evidence omitted from stage) record twelve children, 48 complete chains (1992 pages), and 144 matching measured-first-page associations. Run03 exited 0 in 161.209496 seconds; cgroup peak 804,814,848 bytes, zero swap/OOM events, scratch removed. Borrowing saved 21.3–21.5% requested bytes with empty attributes and 11.7–11.8% with eight repeated string attributes. The absolute saving stayed about 4.19 MB, while rich attributes added about 15.9 MB requested bytes and increased plain first-page CPU by 2.27–2.41 times. Peak allocator savings were negligible. This is synthetic, dictionary-friendly data.

Observer ratios missed their registered tolerance, although all 48 inclusive residuals reconciled. Phase attribution remains inconclusive; there is no isolated attribute-parser or repeatable speedup claim. Compile failures01/02 were preserved, repaired through separately registered supplements, and cleaned with byte-verified archives. Neither contains native performance measurements.

Query Q4

Findings describe optional bounded storage descriptor snapshots. Final-frozen controls02 ran five selected tests, exit0, including publication movement, retention, cancellation and oversize fallback. Controls01 also exited0 but had overlapping source edits; it is not the final freeze.

Frozen builds, preflight, plain pair and counted diagnostic all exited0 with scratch removed. Six children yielded 48 exact complete chains and 36 rejected negative controls. Main workload: 64 published Segments, 8192 seeded logs, four local durable appends and four timed first-page queries at 1 query/second. Plain reuse CPU/wall ratios were 0.968436/0.990878; counted ratios were 1.026921/1.047606. One pair is a screen, with no nomination. Total timing includes append ACK, query serialization and bookkeeping; it is not Spool acceptance throughput. Structural and estimated-byte limits do not establish application RSS bounds.

Disposition

O6/O7 do not clear their correctness dependencies. C5 worker/run-size and O8 collection overlap remain deferred, not silently attempted. No oracle, contract, acceptance threshold, default or release qualification changed. The initial admission and all supplemental protocols were committed separately before their corresponding fresh runs. Frozen evidence retains source identities; later formatting does not replace historical source snapshots.

Verification and cleanup

Final integration, manual documentation checks and resource audit are recorded below after completion. Each coordinator receipt preserves the exact command, exit, cgroup counters, storage location and cleanup state. Builds and grading are included in job peaks; these peaks are not exclusive server RSS. Persistent results live under this repository's data directory; transient fixtures and build caches are on $FABRIC_DATA. Shared build caches are retained; owned finished-job scratch is removed and failure evidence preserved.

catalog-coupled-fast-01 ran cargo xtask checks --profile fast, exit1 in 175.939219 seconds. Workspace tests failed in the unchanged completion_storage missing-table case; Clippy separately rejected a redundant Iterator bound in the new snapshot wrapper. Formatting, compilation, layers, purity, independent oracle/tool controls, properties, fuzz-corpus and simulation checks returned0. The bound was simplified without suppressing the warning; workspace Clippy then exited0 as catalog-coupled-clippy-02. Frozen performance binaries still refer to their original captured source, not this verification-only correction.

Selected descriptor controls03 exited0 (five tests, 9.026434 seconds) after the Clippy correction. catalog-coupled-fast-cleanup-01 exited0 and byte-verified the failed fast-run archive before removing scratch.

catalog-coupled-audit-01 exited0 in 6.007299 seconds with overage/containment/ archive/alias controls. Its final admission inventory recorded 1,769,345,024 allocated aggregate bytes (about1.648GiB), within unchanged2GiB and the192MiB incremental envelope. The mandatory capacity subset plus new failure archives was847,081,472allocated bytes, within its prospective832MiB allocation. This subset is not a certification of all capacity-attributable overhead. Earlier historical capacity excess remains9.54MiB allocated; the current subset against the obsolete768MiB cap includes newly admitted evidence and is not that history.

Conservative all-overhead estimates could not certify either per-lab cap. Query upper bound1,083,404,288bytes exceeds1GiB by9,662,464bytes; this is an attribution uncertainty, not an established exclusive-query overage. Resolve attribution before admitting further workload; aggregate admission does not certify each lab. The audit recorded9698.177383frontier seconds before its own completion, within the existing14400-second frontier. Its sole active receipt was itself. The historical maximum catalog whole-job peak remains about9.08GiB; no project workload exceeded the20GiB/no-swap cgroup envelope. Final receipts and this report add small overhead after the inventory snapshot, reconciled at closeout.

catalog-coupled-docs-01 ran the manual documentation profile, exit0 in 16.023809 seconds: link/syntax checks, checker probes and hook tests passed. Automatic Bun hooks/CI remain disabled. Twenty-five coupled coordinator jobs have terminal receipts, with no active coupled workload at closeout. The largest coupled job was fast verification:8,753,053,696bytes (about8.15GiB), no swap. All newly owned failed scratch was byte-verified into archives and removed; older unrelated failure evidence and shared caches were preserved.