# HP llfix physical kernel/Onload continuation 2026-10-06. This follows the [reviewed generation 3 preparation](RECORD_2026-10-04_llfix-hp-onload-preparation.md) and its [separate physical TX profile](../../bench/llfix_reference/PROFILE_ONLOAD.md). The workload is local construction/send through bundled llfix 1.0.7, with outgoing serialization, one million measured calls and 1,000 warmups per fresh pair. ## Installation and exact inputs The user installed the reviewed payload on HP. Actual file bytes, root ownership and modes match the generation 3 manifest: native binaries/launcher are 0755, manifest is 0444. Manifest SHA256 is `7862839bcfeb7d280e73cc593382e1c155c267f270093bc3c44732fdb46987b5`. Payload `/usr/local/libexec/llfix-reference-hp-v1-d3cb22233f28` and launcher `/usr/local/libexec/libhft-bench/llfix-reference-hp-v1-d3cb22233f28-run` are verified before execution. The existing namespace rule permits their launch. The missing disposable build inputs were restored from the retained archive `hp-onload-preparation3.tar.gz` (SHA256 `e00a57ee4fbacc7eb1f3668836fcf6e95caebc0b0d9c11f0e0caac8408a1a6d0`). All 37 frozen controller/support inputs and 204 original upstream source files match their sealed maps. The original and framework native binaries are the previously prepared GCC 11.5 generation at upstream commit `165cc999fa283a4ead17a7b935c5d3f850fec851`. HP uses its Gold 6254 and SFC9220 physical DAC, namespaces sfcA/sfcB, native CPUs 17/16. Controller and monitoring run on 0-7. Kernel and Onload use the same installed binary bytes, endpoints and persistence policy. Onload's default profile disables application handlers and both TCP loopback acceleration options. ## Retained observer failures and correction Original `functional1` completes its kernel pair, then refuses Onload socket admission. Separate `socket-format1` repeats that refusal and captures the actual stack table. HP's runtime creates **unnamed** stacks, while the original root observer expects `llr-`. Those attempts remain failed; their native summaries do not become admitted results. The user-side observer checkpoint `53c4b7975da582499122fa70985bf37ecf1bc3aa` selects an unambiguous actual native PID in the stack table, including comma separated process lists. It checks owner, executable, namespace and process start time before and after the stack inspection; verifies the mapped selected Onload library; and retains the established physical tuple. It runs as the benchmark user. The installed launcher continues to drop privileges before native execution. A fresh 39-file source capsule has SHA256 `e4d70eef5478e961c5871731f47a71cba66b6ac52ff0837401f0b132b89212be`. The controller explicitly checks the preparation's original frozen source root and independently binds its current controller/observer source closure in each capture and audit. Timing proof must match that observer generation as well as the installed payload/preparation. Four new stack/identity refusal checks and the original ten cold checks pass locally and on HP, zero skips. ## Functional diagnostic capture `functional-observer1` completes all six fresh-process pairs at high priority in the isol pool. Live socket inspections occur during this functional attempt. Each pair has positive physical TX/RX counter deltas on both ports and both native processes return zero. | TX variant | Kernel p50 ns | Kernel p99 ns | Onload p50 ns | Onload p99 ns | | --- | ---: | ---: | ---: | ---: | | Original upstream | 2,412 | 7,133 | 648 | 18,111 | | Framework upstream TSC | 2,781 | 6,583 | 671 | 18,824 | | Framework MONOTONIC_RAW | 2,490 | 7,266 | 679 | 19,522 | These are single functional pairs, with live inspection and concurrent shared work permitted. Their latency is diagnostic. The original first-CPU-MHz scales also differ: 3,099,000,000, 3,100,000,000 and 3,101,000,000 Hz were supplied. No scaling correction is applied to the original output. The complete cold audit **passes**: 12,012,000 persisted orders, required converted values, both parties, contiguous owners, FIX framing/checksums and four million framework uint64/HFTBRW samples, including quantiles and actual maxima. The functional manager completes all four high-priority steps with zero return codes. Its grade is `physical-functional-wire-raw-pass`, with timing admission false. ## Timed capture and rejected activity review The separate quiet manager binds the passing functional proof and takes a high-priority isol-no-shared lease. Its existing 120-second preflight passes: no active-process violations, no load violations, correct isolated native CPUs and disjoint controller placement. All 18 fresh pairs complete with both native processes returning zero. Transport order reverses in the second round. The continuous activity review **rejects** timing. It covers all 18 pair windows with 86 sampled seconds and observes kswapd0 up to 24%, filesystem writeback/event workers up to 14%, kcompactd0 at 5% and fprintd at 6%. The unchanged 5% threshold includes kernel/housekeeping work; no ownership exclusions are inferred. The independent BEELINK2 review uses the exact frozen auditor and capture/pidstat bytes and returns `external-activity-observed` with exit code 1. A prior shell wrapper returned the final `cat` command's zero exit; its rejection JSON is preserved and the separate direct CLI continuation retains the actual nonzero exit. The following summaries are **diagnostic median-of-three**. The complete timed capture's wire/raw audit passes all 18 pairs: 36,036,000 persisted orders and 12 million framework uint64/HFTBRW samples. Required values/groups, contiguous owners, framing/checksums, quantiles and actual maxima pass. The activity rejection prevents quiet timing admission independently of that audit. | TX variant | Kernel p50 ns | Kernel p99 ns | Onload p50 ns | Onload p99 ns | | --- | ---: | ---: | ---: | ---: | | Original upstream | 2,641 | 7,052 | 648 | 18,406 | | Framework upstream TSC | 2,821 | 7,018 | 669 | 18,336 | | Framework MONOTONIC_RAW | 2,820 | 7,225 | 680 | 19,069 | The current [vendor TX/send baseline](https://www.llfix.net/), checked 2026-10-06, reports Onload p50 573ns and p99 2,663ns on Gold 6134/SFN8522 PLUS/RHEL9.4/GCC11.4.1. Its displayed fixture uses nanosecond timestamps; this pinned original 1.0.7 TX fixture uses milliseconds. The original HP median is in the same range, while its p99 is approximately 6.9 times the reported value. Vendor reproduction is false; this default-profile HP generation does not qualify that baseline, the separate RX workload, Stage 2, full-session admission or release. The actual HP activity CLI also returns 1 with `external-activity-observed`, matching the independent BEELINK2 review. The quiet manager retains its original failure and does not generate a positive timing review. Upstream clock-scale equality is false: 3,099,000,000, 3,100,000,000, 3,101,000,000 and 3,103,000,000 Hz occur across the upstream arms. The MONOTONIC_RAW samples retain their independent clock label. ## Sealed evidence and publication HP retained the completed native captures, journals, raw samples, exact original and controller sources, preparation, installed-input manifests, failed observer attempts and manager receipts. The archive contains 854 sealed files and is 938,371,792 bytes compressed. Independent streaming verification on HP and BEELINK2 checks every member's SHA256, size and mode, rejects unsafe/duplicate members, and returns identical `replica-verified` receipts with timing admission false. Archive SHA256: `78943b1749217d12bb4911c3b70d12ced1ed312aa6be90d7bbc9fab90b4d26ae`. Seal SHA256: `41479192ace42f1f2881d344dde59713d6b1d09c17891e86804b7b46e38d7ae1`. The exact compressed replica is retained on both machines at `~/archive/libhft/codex-llfix-reference-20261006/hp-onload1.tar.gz`. The installed payload and frozen preparation remain available for a separately pinned future generation. The passing captures comprise 24 pairs, 48,048,000 persisted orders and 16 million framework raw samples; those counts exclude the earlier failed attempts. The [final selection](evidence/llfix-hp-onload-20261006/final-selection-manifest.json) binds 19 exact metadata files, including the [18-pair cold audit](evidence/llfix-hp-onload-20261006/onload-generation3/captures/quiet-observer1/audit.json), [activity refusal](evidence/llfix-hp-onload-20261006/onload-generation3/captures/quiet-observer1/activity.json), [HP archive receipt](evidence/llfix-hp-onload-20261006/hp-onload-retention1.json), [BEELINK2 archive receipt](evidence/llfix-hp-onload-20261006/hp-onload-beelink2-replica1.json) and [full seal](evidence/llfix-hp-onload-20261006/retention/hp-onload1/seal.json). Operational retention/replication receipts generated after sealing are explicitly outside the archive seal. The earlier 191-file selection remains unchanged as its original checkpoint. All 39 frozen controller/support input bytes and modes remain unchanged at integration; no new SDK/native build is claimed.