# HP bounded-retention timing continuation, 2026-10-12

This fresh `capture2` follows the retained
[zero-pair quiet refusal](RECORD_2026-10-12_hp-llfix-bounded-timing.md).
The source and 29-control qualification remain pinned at
`35d637a1a335148ab08ec07f91a22db6b83cc8c6`; its 15-file capsule SHA256 is
`48ab8b13dda133931c0c197566265c03b0f2e1c067df20272755cef1a85c97fc`.
No native binary or source is rebuilt or modified. This continuation reuses the
passing three-pair bounded functional proof only after exact source, installed
SDK/runtime, prepared inputs and current host policy verification.

The functional capture SHA256 is
`1b3260edcd38c2d4ca2a2fdaf7137190f350c7eca273bb403d28b07d5bbec58a`;
its cold audit SHA256 is
`96d69f0e33b2d4ad14b9f8bd54249a6275b70e917517abd36e7672a23cad705f`.
The original 105-input archive/seal remain immutable at
`f359966c862c904706b405e333df52c20ec89a8924cfbc86b1c0696520ae2510` /
`c42d910546444d84adb80a096e07b2fe84a04efd2fda06cd07b8a6207ade2cbf`.
The old failed attempt is not overwritten or regraded.

## Fresh readiness and declared work

Read-only HP status found both CPU pools free and irqbalance inactive/dead,
PID 0. A fresh 30.015-second observation under high-priority isol-no-shared
found no process at or above 5%. It validates all 15 frozen source inputs and
retains the original activity rows; it does not substitute for the full quiet
preflight or certify continued quietness. No process was paused or stopped.

`capture2` uses the unchanged explicit bounded timing method: three fresh pairs
per arm in forward/reverse/forward order, each 1,000 warmup and 1,000,000 measured
orders. Each pair requires the full 120-second preflight, followed by actual
native TX/receiving-peer work, a three-second monitored tail, then two complete
verified lossless pair archives before raw pruning. Full native persistence
remains unchanged. Before the next pair the full quiet preflight repeats and
host policy must match. Observers stay on 0–7; native client/server use 17/16,
Solarflare Onload DAC and the frozen original llfix 1.0.7/GCC 11.5 binaries.
Receiver-only spin and 1 MiB buffer remain the protected rx16 profile.

Immutable operations16 use HP high-priority shared verification/audits/retention
and isol-no-shared capture. Another owner's job is neither stopped nor changed.
No SCAN contact, helper pause, OS/service/IRQ/turbo/sysctl/cache-policy write,
SDK/root runtime, shared API/framework/clock, Stage 2 or full-session edit occurs.
Current framework preallocation and Stage 2 llfix 1.0.8 remain separate epochs.

Durable worker PID 2740777/start ticks 208249852 launched 20:36:50 UTC.
Frozen source, installed runtime, prepared inputs, exact functional proof and
current host policy verification passed immediately. The full preflight then
ran 20:36:50–20:38:51 UTC (120.01 seconds) and rejected another process using
at least 5% of one CPU. **Zero timed pairs started.**

| Observed process | PID | Peak CPU |
| --- | ---: | ---: |
| git | 2741604 | 100.3% |
| kworker/u72:3+flush-253:2 | 2727595 | 19.7% |
| kswapd0 | 183 | 7.9% |
| kworker/u72:2-flush-253:2 | 2740632 | 7.9% |
| kcompactd0 | 161 | 6.9% |
| codex | 892412 | 6.9% |
| kworker/3:2-xfs-conv/dm-2 | 2740284 | 5.9% |

The Git identity's retained context reports cwd
`/home/yann/code/wt/codex-review-wire-1012/libhft`, PID start ticks 208255813,
parent 2741602 and process group/session 2741503. This identifies a concurrent
repository operation, not permission to stop it. Code was below the threshold
in this attempt. Load1 changed from 0.00 to 0.11 and was not the blocker.
Kernel observations identify pressure/writeback activity but do not establish
that Git caused it.

The first source generation's rejection and this new original capture/preflight
remain intact. Capture cold grading was not run without a completed capture;
no numeric comparison exists. The reused functional proof remains valid at its
original pin, and no additional native process was executed.

The short readiness observation was insufficient to establish continued
quietness. The portable microbench guide now explicitly includes repository
operations, source staging, transfers and cold grading in the whole-host quiet
reservation. This documentation changes no threshold, qualifier, runtime,
workload, clock or historical grade. Actual continued quietness is still required
for the nine-pair diagnostic. Native summaries alone are
not quiet timing, clock-parity, vendor, SDK or report admission. Any new refusal
retains its original capture, source/operations and complete raw evidence.


## Exact retention

The complete 103-input continuation archive verifies on HP and BEELINK2,
113,798,153 bytes, at
`~/archive/libhft/codex-llfix-reference-20261012/hp-framework-bounded-timing2.tar.gz`.
SHA256 is `196085e41d21e273d9541c93ba25ffca876f5d3bc844afb437f6ac4cb264a748`;
seal SHA256 is `741fed7218a3f337e71abe02b62e03db41cf7ef5e9914235f1b634032357b6e8`.
The unchanged verifier SHA256 is
`86f05bac518be7662b7b3f506880bf443f4bb09100bf9f099a64df5c0eb3b56a`.
All original file bytes, sizes and modes match, including the reused full
functional proof's lossless pair archives. The old failed capture1 remains
separately sealed and is referenced by immutable input hashes.

The [compact selection](evidence/hp-llfix-bounded-timing2-20261012/selected-files.json)
contains 103 original files/1,858,235 bytes plus its index; the eight excluded
source/pair archive containers remain in the full archive. Git preserves the
selected bytes and executable flags; original full modes remain in metadata
and the lossless archives. No journal or sample is downsampled.

A whole-host quiet window, covering non-benchmark repository/transfer work too,
is needed before the next fresh attempt. This continuation changes no source,
SDK, installed ELF, activity or clock rule, and makes no timing promotion.