# HP capture-v1 quiet timing controller and compact output

2026-10-17. Source/controller qualification only; no timing result is published.

The [new timing lane](../../bench/llfix_reference/PROFILE_CAPTURE_TIMING.md)
reuses the [qualified physical capture-v1 generation](RECORD_2026-10-17_hp-llfix-capture-physical.md).
It implements six fresh-process measurements, three per clock, with a complete
unchanged120-second/5% host guard before every pair and continuous one-second
pidstat coverage around measurement. Only a completed, independently cold-graded
campaign can produce the two compact result keys.

## Functional qualification

Controller source `55778fb4e6ffb1f895e44cb0eedcc593491182e3`, retained on
`s/codex-llfix-capture-timing-source-20261017-v2`, completes an HP high-priority
`isol` functional lease, with concurrent shared work allowed:

| Check | Actual outcome |
|---|---:|
| Focused checks, local and HP | 5 pass, zero skips |
| Physical Onload clock pairs | 2 |
| Measured samples checked | 514 |
| Checked orders per peer, including capture/warmup | 568 |
| Native exits | 4 zero |
| Pidstat sample timestamps per pair | 6 |
| Unowned activity observations per pair | 2 |
| Compact timing publication | refused |

Both physical pairs pass actual body-prefix, converted-value/owner, raw-series,
process, affinity, mapped Onload/socket and counter checks. Cold replay agrees.
The live machine-profile audit passes and observes72 Mellanox/Solarflare IRQs.
Loaded irqbalance stays inactive/dead/PID0; policy snapshots agree before/after.
No SDK, root helper, service, sysctl, IRQ or turbo setting is changed.

Pidstat correctly detects concurrent external activity. Functional mode neither
creates a quiet pass nor admits these samples as timing. Its explicit attempt
to publish them refuses before a result directory is created. All eight
selected controller/support inputs retain their bytes/modes after execution.

## Final source binding

Source `9ead704615634c7d9f8f09261ea7cebc0f66299c`, retained on
`s/codex-llfix-capture-timing-source-20261017-v3`, adds one further admission
requirement: the completed functional proof must name this exact controller
source pin and all eight current source/support hashes. It must contain both
actual clock pairs and the publication refusal. That proof hash travels with
each future result. An actual read-only check against the completed v2 receipt
refuses its older identity; a previous generation's pass cannot qualify v3.

Six focused checks pass locally without skips. The final v3 HP functional
generation is queued high-priority behind other owners' existing exclusive
work and shared conformance. Their processes and queue entries are unchanged.
The v2 physical pass is not relabelled as a v3 pass. No timing job is queued.

## Publication and controls

The compact writer pools three million full-width samples per clock. The actual
serializer is tested with an inert three-million-sample vector containing one
rare maximum: p99 remains10ns, the maximum123456789ns survives, the CDF ends at
that maximum and has at most2000 points, and the file is below256KiB. These
numbers describe a numerical test fixture, not engine performance.

Negative controls refuse shortened/noisy/NaN quiet evidence, changed affinity,
excess load, incomplete/reused/failed pairs, changed captures, missing physical
proof, missing/changed functional provenance and functional-to-timing promotion.
A refused mocked preflight retains its failure receipt without a native launch.
No real quiet preflight was attempted in this sitting.

The result identifies the local construction/send boundary explicitly: no
execution report is awaited. `closed` names the local call loop. The upstream
clock retains and discloses its original integer CPU-MHz divisor; the other
uses the frozen `CLOCK_MONOTONIC_RAW` implementation. No historical number is
rescaled or promoted. Actual PCI identity is SFC9220/[1924:0a03] on both ports;
firmware is explicitly unobserved through the fixed namespace facts interface.

The [compact result contract](../../../docs/bench/BENCH_RESULTS.md) governs
future output. Raw logs/samples and controller working data remain outside Git.
The19 older native and14 physical source inputs, their binaries/SDKs, installed
data generations, shared framework/clocks and Stage2 files are unchanged.
Only HP was contacted. A coordinated whole-host quiet window and the final
matching functional proof remain required before actual timing.

## Later qualification in this sitting

The [continuation](RECORD_2026-10-17_llfix-capture-common-toolchain.md) records
the completed exact HP v3 controller and candidate-v1 functional qualifications.
The pending status above describes this original implementation checkpoint;
no timing or physical candidate grade is inferred.