Historical report imported 2026-10-10 from docs/experiments/benchmarks/performance-refinement-fedora-run-01.md. Original SHA-256: 018ea50b5a5bcb499dc93e7184556eff3692bc04be3b49d53ed44ac876871e47. Measurements and outcomes retain their stated revisions. Links and local environment labels are normalized; $FABRIC_DATA is the mounted experiment drive. See evidence policy.

Fedora performance refinement, run 01

In progress under the separately committed protocol (local checkout reference docs/experiments/benchmarks/performance-refinement-fedora-protocol.md; not present at origin/main), on milestone/streaming-segment-output containing a6d4905. The cloud isolation screen remains completed finite evidence; it is not rerun or called qualification. No installed defaults or durability/custody rules change.

Host and rate correction

Environment capture (local evidence omitted from stage): Fedora 44, i7-10750H, 6 physical cores / 12 SMT CPUs, affinity 0–11, 33,491,931,136 bytes RAM, 8 GiB unused compressed swap at capture, about 25.3 GB available RAM. Every inspected ancestor has unlimited CPU/memory high/max (except the root, which exposes no such files); process open-file limit 1,024. Repository storage is LUKS-encrypted NVMe/Btrfs with about 494 GB free; /tmp is tmpfs. Rust/Cargo 1.99, Python 3.14.7; Bun absent initially, native Bun 1.4.2 downloaded into owned /tmp for documentation checks. The cloud had five affinity CPUs, four quota CPU equivalents, shared 16 GiB memory and buffered cloud storage, Rust 1.98. This workstation has background desktop processes and thermal/frequency variability; it is neither a dedicated device benchmark nor the Debian deployment profile.

The unchanged retained diagnostic initially could not build offline (missing local Cargo cache, exit 101 inside its subprocess); cargo fetch --locked completed with exit 0. A new diagnostic invocation then exited 1 with all three original disagreements reproduced: output (local evidence omitted from stage). The original audit inputs, source, receipts and expected outputs are unchanged.

The repair caches existing display sort keys once, uses structural map equality for series identity, and breaks ambiguous display-key ties structurally. It retains time/identity/sequence/index order. A typed pure-core adapter computes integer deltas in i128 and compares mixed represented values without premature integer rounding, then delegates time/reset/finite-rate decisions to the existing kernel. Mixed subtraction keeps the independent oracle's floating arithmetic; it does not invent precision absent from an OTLP double. No dependency was added to the core and the Python oracles were not changed.

The HTTP integration regression covers the three retained failures, multi-point colliding maps, mixed numeric pairs/fractions/decreases, the full i64 range and 2,048 independent series. The corrected implementation produced 2,060 rows; its targeted Cargo test exited 0 through durable TLS delivery, scan/walk, journal, sealed Segment and restart, with independent oracle grading. Restoring the exact original query source made this same regression exit 101, with 2,062 rows (invented cross-series rates): negative-control receipt (local evidence omitted from stage) and failure (local evidence omitted from stage). This is independent behavior evidence, not a claim about arbitrary counter inputs.

Initial fast profile exited 1: 19 checks passed, Clippy failed on the previously recorded Rust chunk-iterator lint. After replacing the two existing iterator instances with typed as_chunks, the second profile also exited 1 solely on an unnecessary borrow in the new test; that borrow was corrected. Both failed profiles and their receipts are retained. The direct final Clippy command exited 0: commands (local evidence omitted from stage). Semantic-mutant outcomes and subsequent attribution/composites are pending. These records do not claim a full fast-profile pass.

Continuation status

The registered next measurements add native collection/OTLP, lifetime and builder/query phase attribution, then independent journal granularity, worker counts including three, selected interactions and sequential native consumer composites. None is claimed complete by this correctness record. Default-sized files and memory caps remain hypotheses; enterprise and QUIC remain secondary.