# HP llfix bounded timing capture4, 2026-10-13

This fresh continuation follows the retained
[capture3 VS Code refusal](RECORD_2026-10-13_hp-llfix-bounded-timing3.md).
Source remains `35d637a1a335148ab08ec07f91a22db6b83cc8c6`, with 15-file
capsule SHA256
`48ab8b13dda133931c0c197566265c03b0f2e1c067df20272755cef1a85c97fc`.
Native binaries, SDK, preparation and the earlier three-pair functional proof
remain pinned. No source or runtime rebuild occurs.

## Declared fresh attempt

At 17:22 UTC both HP pools are free, irqbalance is inactive, and no compiler,
JVM, CLR or Git process appears in the read-only status. The same VS Code
renderer remains present. Its lifetime CPU average of 0% is not a quietness
proof or confirmation that the requested quiet window has been reserved.
After the nine-hour interval this attempt goes directly to the full unchanged
120-second activity preflight, without another short readiness observation.
No application, service or other owner's job is stopped or changed.

Operations19 launch durable worker PID 2983821/start ticks 215743599 at
17:25:47 UTC. Verification, cold audit and retention use HP shared/high;
timing uses isol-no-shared/high. Controller/observers stay on CPUs 0–7,
native client/server on 17/16. The declared matrix remains nine fresh pairs:
three per original/framework-upstream/framework-monotonic arm, in balanced
forward/reverse/forward order, with 1,000 warmup and 1,000,000 measured orders.
Every pair repeats the full preflight and policy check, retains a three-second
monitored tail and two verified lossless pair archives before raw pruning.

The original llfix 1.0.7/GCC 11.5 native generation, full persistence, Solarflare
Onload DAC and protected receiver-only spin/1 MiB buffer profile are unchanged.
The reused functional capture and cold audit hashes remain
`1b3260edcd38c2d4ca2a2fdaf7137190f350c7eca273bb403d28b07d5bbec58a` and
`96d69f0e33b2d4ad14b9f8bd54249a6275b70e917517abd36e7672a23cad705f`.
All original failures and full functional archives stay preserved. This new
archive retains exact proof bindings instead of another staged raw duplicate.

No SCAN contact, helper pause, threshold relaxation, OS/service/IRQ/turbo/sysctl/
cache change, Stage 2/shared framework/clock/helper, SDK or full-session source
edit occurs. Current framework preallocation and Stage 2 llfix 1.0.8 remain
separate generations. A successful native summary cannot establish quiet
timing, clock parity or the vendor's different hardware/transport numbers.

## Actual outcome and retention

Frozen source/runtime/prepared inputs, reused functional proof and current
host-policy checks pass. The full preflight runs 17:25:48–17:27:48 UTC
(120.01 seconds), then refuses unrelated activity above the unchanged 5% gate.

| Process | PID | Peak CPU |
| --- | ---: | ---: |
| git | 2984086 | 98.3% |
| kworker/u72:2-events_unbound | 2982466 | 21.6% |
| codex | 892412 | 19.7% |
| kworker/u72:1-events_unbound | 2981652 | 10.8% |
| kworker/u72:3-events_unbound | 2973412 | 5.9% |

The Git identity retains start ticks 215749613, parent/process group/session
2983969 and cwd `/home/yann/code/wt/codex-libhft-hp/libhft`, observed at
17:26:49 UTC. This identifies another session's repository work; it is not
permission to stop it. Codex identity/context and the original permission
refusals for kernel cwd reads remain retained too. VS Code is not an observed
threshold blocker in this attempt. None of these rows establishes that Git
caused the kernel activity.

**Zero native pairs started.** Original controller/worker failures remain
intact. Capture cold grading and numerical comparison are not run without a
completed matrix. Retention succeeds; the worker exits. The old three-pair
functional proof remains valid only at its own original pin.

Free CPU pools and a quiet lifetime CPU average again failed to guarantee a
quiet interval. The next run needs coordinated quietness across application,
repository and background work; repeating the same unchanged reservation
without coordination is not a new performance investigation.

## Complete archive and committed evidence

Both HP and BEELINK2 completely verify `hp-framework-bounded-timing4.tar.gz`:
35 sealed inputs, 276,948 bytes, every original byte/size/mode preserved.
Archive SHA256:
`14496a9c8311fa8c34573307fe5c5c1b5fc8a67b28609303b9c60b0b095f04ce`.
Seal SHA256:
`0d549db4ce4755cb436d2c4a76e221522c2b6da94558f0510304eea95b738d34`.
Both archive roots are `~/archive/libhft/codex-llfix-reference-20261013/`.
The unchanged full streaming verifier SHA256 remains
`86f05bac518be7662b7b3f506880bf443f4bb09100bf9f099a64df5c0eb3b56a`.

The [selection index](evidence/hp-llfix-bounded-timing4-20261013/selected-files.json)
binds 42 files/574,263 bytes plus its index. Only the source tar container is
excluded from the compact selection; it remains in both full archives.
Original full modes remain in the archives/index, while Git preserves exact
bytes and executable flags. SSH logs retain their original carriage-return
bytes. The prior complete functional proof is referenced through immutable
bindings instead of another duplicate raw staging copy.

After normal main integration, independent cold checks match all 42 selected
Git blobs/executable flags and all 15 frozen source hashes. They reproduce
the five actual blockers, exact Git PID/start/cwd identity, successful input
verification and retention, original failed worker status, and zero native
pairs. No capture audit or matrix output exists to promote.

No old receipt or grade is overwritten, no raw evidence is downsampled, and
no timing, clock-parity, vendor, SDK or report promotion is claimed.