Historical snapshot: the canonical repository body is retained because a harness consumes it. Keep edits to that versioned input in the repository; this wiki copy records its migration snapshot.

Historical report imported 2026-10-10 from docs/experiments/benchmarks/cross-system-continuation-findings.md. Original SHA-256: 5ce22f6cfefaab1ef8dd292d1254fe34f9d10554e0b2e054bb719e54d8decb23. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Cross-system dissection and diagnostic census

The registered continuation (local checkout reference docs/experiments/benchmarks/cross-system-continuation-protocol.md; not present at origin/main) downloaded and dissected the remaining supplied FOSS, then ran bounded Fabric comparisons. The source synthesis links the three detailed dissections. Source mechanisms are hypotheses for Fabric; no upstream implementation was benchmarked or adopted as a dependency.

Source retrieval and evidence integrity

Twenty-one full archives were downloaded and their regular members hashed. ClickHouse exceeded the 128 MiB compressed cap; Vector and FoundationDB's older pins returned HTTP 404. Selected-file retrieval covered all three, using explicit new immutable pins for the latter two. Supplement 02 includes actual FoundationDB simulated persistence and Vector disk-buffer recovery paths; earlier wrong-path 404s remain recorded. Turso's earlier successful retrieval is separate.

All 24 retrieval records and both supplementary archives passed the independent evidence checks (local evidence omitted from stage), exit 0. This validates retained artifacts and recorded failure prefixes, not successful complete retrieval of the three failed archives. Rejection controls covered altered, missing, duplicate, unsafe and linked archive members, invalid pins, interrupted downloads and oversized response prefixes.

The review found an interruption-preservation defect in the new supplementary fetcher: it removed completed temporary responses before the aggregate archive existed. Responses and per-file receipts now stay on disk until exact aggregate readback; the deterministic interruption control preserves the failed prefix. Historical successful artifacts were rechecked and remain unchanged.

Memory: physical layout and observer ownership

memory/census-01 completed six native children but failed its wide-row metadata comparison: reference filter groups 1, bounded groups 3, with identical 1024 log rows and ordered custody. FileEntry.rows means physical group count for a filter, logical row count for Parquet. ADR-0022 permits the byte cap to close a group early. The failed run remains failed, exit 1.

A separately registered checker correction authenticates files, decodes each Parquet layout with pinned PyArrow 22.0.0, reconstructs exact FTF1 Bloom bytes from its row strings in Python, and requires own-layout alignment before comparing logical metadata. Missing, misaligned and corrupted filters must be rejected. This adds independent filter evidence; the ordered row ledgers still use the same Rust reader and are not a full independent query oracle.

Census 02 (local evidence omitted from stage) passed all six cells and these checks. Single-pair build-only measurements at 16 MiB input:

Shape	Reference requested-live peak	Bounded requested-live peak	Instrumented build time, reference → bounded
Steady	98.99 MiB	34.33 MiB	346.1 → 369.3 ms
Adversarial	99.28 MiB	34.36 MiB	375.6 → 393.0 ms
Wide rows	95.04 MiB	32.06 MiB	262.4 → 260.0 ms

These show 65–66% less requested-live build peak on these small fixtures; they do not replace historical results at different sizes or establish a service memory ceiling. Reservations are not observed. Native CPU/wall and Python validation time are recorded separately.

Census 02's parent imported PyArrow before spawning native cells. Its bounded wait4 RSS rose to 93–106 MiB while post-exec sampled high-water remained around 51–57 MiB. This confounds a native-only peak interpretation. The registered repair puts decoding in a separately reaped child, records spawning-parent RSS, and reuses old archives only after exact path/size/digest equality. Historical RSS values remain labeled; they are not silently corrected.

Census 03 (local evidence omitted from stage) completed the repair and all six exact archive reuses, with unchanged native binary/fixtures:

Shape	Reference whole-worker RSS	Bounded whole-worker RSS	Reduction
Steady	112.21 MiB	54.23 MiB	51.7%
Adversarial	112.12 MiB	51.53 MiB	54.0%
Wide rows	105.92 MiB	57.20 MiB	46.0%

The spawning parent stayed below 35 MiB without PyArrow imported. Post-exec sampled high-water and wait4 peaks agreed within about 0.33 MiB, supporting the observer-contamination explanation for census 02. Whole-worker RSS includes fixture generation and native validation; it is not build-only memory. These remain one pair per shape, and /proc sampling can miss short peaks. Build times in this replay changed +9.2%, +2.8% and −4.9%; no universal speedup follows. The 405.3 MiB whole-job cgroup peak includes the isolated decoder and cache.

Query: full-chain evidence and remaining block work

Query census 01 (local evidence omitted from stage) completed eight plain/counted variants: 512 complete chains, 3520 pages, and 72 rejected negative-control verdicts. All source and raw-data archive readbacks completed. The unchanged oracle grades every continuation, not just the first page.

At 2048 rows with 1024-byte bodies, counted sealed selective queries requested about 11.94 MB despite returning one row. Parquet loading/projection accounted for about 92% of their counted query CPU. Scan and Walk were similar on these selective fixtures; existing pruning strongly helped the empty-tail case. Counted/plain timing differences also changed fault and system-CPU costs, so they do not isolate observer overhead. Nested spans are not additive. Pure plain full-chain query timing, controlled cold caches and service QPS remain unmeasured.

The existing borrowed-log path is the next small-profile ablation. Its earlier CR2 experiment already saved allocation at 65536 rows but failed CPU/wall guards and was not nominated. The new 128/2048-row screen keeps that rejection, existing corruption controls and unchanged oracles. It cannot nominate production using an easier allocation-only criterion.

The small-profile ablation (local evidence omitted from stage) completed 16 variants, 1024 accepted complete chains, 7040 pages and 144 rejected controls, exit 0. All eight registered selective wide-body calls saved 2,110,304–2,110,340 requested bytes, meeting the at-least-1-MiB prediction. This is allocation churn, not live peak or RSS. Sealed selective plain repeated CPU medians changed +2.15% for Scan and +0.06% for Walk; their wall changes were +1.01% and −0.64%. Other sealed populations were slower in this sample, up to 13.3% CPU. An already-borrowed tail control also varied by 18.5% CPU, limiting causal interpretation of the timing differences.

Each cell has one fresh process per arm and three repeated calls; balanced order across cells does not create independent within-cell replications. The allocation mechanism reproduces at smaller scale, while a compelling CPU benefit does not. CR2's no-nomination remains; no default changed. The useful next discriminating question is avoided block decoding and its integrity/conversion costs, rather than another claim that string-copy removal alone will accelerate queries.

Operations: useful mechanism, inconclusive universal gain

The native Store/intake comparison uses the actual candidate sealer and an example-local historical group-join baseline. Every case has 128 stream identities, three initial journals and 96 live offers. Shapes are balanced three workers, skewed three workers and a one-worker mechanism-negative control. The three seed labels perturb body text, not independent scheduling distributions.

prefix-load-03 completed 18 cells but used a 2 ms observer instead of the registered 10 ms. It remains diagnostic. The corrected run 04 (local evidence omitted from stage) completed 18 cells at 10 ms: 144 accepted complete query chains, 36 rejected controls, exact custody/restart retry, and 96/96 accepted Batches in every case with zero Unavailable. No HTTP/TLS or sustained-capacity claim follows.

Only one skewed pair had non-overlapping journal byte-time bounds, with 11.61–42.88% savings. The other skewed pairs and all balanced/one-worker pairs overlapped. Skewed ACK p99 improved 15–20% in run 04, but one one-worker pair worsened 6.3%; the 2 ms diagnostic had different adverse ACK outcomes. With 96 ACKs, nearest-rank p99 equals the maximum. These are finite observations, not stable tail-latency estimates or a universal scheduler speedup.

The 10 ms observation widened byte-time uncertainty and did not resolve early release before the middle builder's publication directly. Polling can also perturb the workload; its interval width is not the only possible cause of different runs. Exact checkpoint service duration remains unmeasured. A future test should timestamp lifecycle events with measured observer cost and vary the offered cadence around the commit thread's quiet interval.

Reproducibility, failures and cleanup

Commands, exits, source/protocol snapshots, cgroup observations and cleanup are under coordinator (local evidence omitted from stage); machine-readable descriptive results are in report.json (local evidence omitted from stage). All project workloads used the resource launcher, 16/20 GiB high/max, no swap, serialized admission and data-drive scratch/cache. No remote server was used. Successful source archives/fixtures were removed only after exact preservation. Failed retrieval prefixes and native failed-case archives remain. The cleanup receipt (local evidence omitted from stage) removed only empty owned failure directories and charged the 3.280-second pre-admission refusal to the continuation ledger.

Other preserved preparation failures were the example's use of a private clock (compile exit 101), redundant nested coordinator lock, and absent output parent (both exit 1 before a native cell). The example now implements the public clock port locally; the runner respects the outer lock and creates its owned parent. No production API was widened and no historical verdict was overwritten.

The finite census is partial M1/Q1 evidence. Mixed service memory reservations, long soaks, persistence-fault campaigns, alternate formats, metrics/traces query census and deployment qualification remain separate work. Verification and the registered diagnostic follow-ups are recorded below as they complete.

cargo xtask checks --profile fast ran through the continuation launcher and passed all 17 gates (local evidence omitted from stage), exit 0, 279.3 seconds outer service time, 575.1 MiB cgroup peak, zero swap/OOM. This includes the workspace and existing rejected-row/corruption controls. Later changes in this round are diagnostic Python drivers and documentation; no subsequent Rust/product change invalidates that receipt.

The first manual documentation attempt returned exit 3, environment unavailable because Bun was not on PATH. That result and all three unavailable receipts are preserved in final-docs-01. Reusing the existing authenticated runtime archive in data-drive scratch, the retry (local evidence omitted from stage) passed all three unchanged manual checks, exit 0. The runtime copy was verified and removed; no hook or CI job was enabled. The resource summary (local evidence omitted from stage) and copied verification receipts retain the current census of commands and limits.

This round's next research priorities are avoided block decoding with exact integrity/fallback costs, completion-owned memory under simultaneous service load, and event-based persistence/reclamation timing. Alternative formats and failure-image campaigns remain unexecuted hypotheses. The source dissection and the registered finite census are complete; the broader research goal is not.