# HP llfix resource observation, 2026-10-12

This fresh diagnostic follows the [rejected quiet5 matrix](RECORD_2026-10-12_hp-framework-tx-quiet5.md).
Quiet5's raw evidence, activity rejection and unequal clock divisors remain
unchanged. No timing, vendor, Stage2, SDK or release admission is claimed.

## Source and method

Quiet5's seal contains15,099,494,400 logical bytes of binary journal segments,
compared with90,063,258 bytes of framework CSV/HFTBRW samples. Thus the large
logical footprint primarily comes from the original persistence workload, not
CDF presentation. Logical file size is not resident RAM or allocated disk usage;
the new footprint witness records allocated bytes separately. The earlier
postrun memory snapshot showed22,998,616kB total RAM, but supplies no original
peak or measured-interval resource deltas.

Private source `2fd1f825d8e4ce89a3403307405ce307f638a288` adds a bounded
read-only observer and explicit controller/pidstat identities. The 13-file
source capsule is `738fe0848bb50573046fe8480b95bd66fcbc32b542b8ddd2935ce0667c0eb197`;
all selected Git bytes and modes are checked before remote preparation. All 17
local checks pass. The matrix function AST and original activity checker bytes
are unchanged from the quiet5 epoch.

Every second and at pair boundaries the observer retains original `/proc/meminfo`,
`/proc/vmstat` and the owning filesystem's block-device counters. Pressure files
are read if available; HP's kernel currently has neither memory nor I/O pressure
file, and absence is recorded explicitly. After each pair, a stat-only traversal
records the attempt's logical/allocated footprint. It never follows symlinks or
reads journal data during native work. Controller/monitor identities include
PID, start ticks, parent/group, UIDs, executable hash and actual CPU affinity.

Cold review reparses original readings and refuses changed parsed values,
incomplete readings, boot/device changes, counter resets and observer failures.
Reported memory extrema are sampled values. Counter deltas describe the whole
host, including shared work, and cannot identify a causal process. The original
global activity gate still includes the controller and kernel workers.

The three functional arms reuse the protected original llfix 1.0.7 TX,
framework RDTSCP and framework MONOTONIC_RAW binaries, original receiving peer,
Onload DAC CPUs17/16 and receiver profile. Each sends1000 warmup and1000000
measured orders. Nothing is rebuilt, installed, tuned or paused. The new attempt
is `onload-load-generation1/captures/hp-framework-resources-20261012-1/functional1`
under this session's HP build root. Original functional/body/owner/raw checks
remain mandatory. This is a functional resource diagnostic with shared work
allowed, not a quiet run.

## Execution

Small-source staging waited shared/high behind another session's exclusive
HP lease. Durable worker2354162/start204330348 launched09:43:35UTC. Preparation
completed09:43:44 (17 checks, zero skips), then all three functional pairs
completed09:44:01 with six native exits0. The unchanged full wire/raw cold audit
started09:44:01 under shared/high and completed09:49:34, exit0. All6,006,000
order copies pass required values, groups, owners, envelope and counts; both
framework arms pass all two million CSV samples against their HFTBRW originals
and independent summary statistics. Lossless retention completes under shared/high.
The functional isol/high lease permitted shared work. The audit budget is one hour; the earlier quiet5
900-second outer timeout is retained and not repeated as this generation's limit.

No SCAN contact or IRQ, governor, turbo, service, sysctl, cache or kernel-thread
control occurs. HP verifies all117 named archive inputs in100,048,958 compressed
bytes. Archive SHA256 is
`7de4ef1e342929e74b72688a8926a977462ef9f416c315ce9599e87c224fc4c3`;
seal SHA256 is
`34d30e64a18e9983efa2abb80c29adeead9444704ca13b9f6a8ed06c09e3b6e1`.
Local full-archive replication waited shared/high behind another exclusive
owner. Complete streaming verification on BEELINK2 agrees with HP on every
named file's bytes, size and mode, plus archive/seal hashes. No downsampling or
lossy rewrite occurs. [The compact selection](evidence/hp-llfix-resource-observation-20261012/selected-files.json)
retains93 source/metadata files (1,670,493bytes);32 original raw/store/container
files remain losslessly in the archives, excluded from Git. Both copies are
under `~/archive/libhft/codex-llfix-reference-20261012/`.

The frozen policy confirms18 online CPUs with singleton sibling sets, including
16/17; client and server did not share an SMT core. No topology change occurs.
The SDK/root payload/source-capsule bytes remain pinned. No native or quiet job
remains; publication is read-only with respect to both retained generations.

## Resource observations

The independently parsed resource stream has21 samples over13.704seconds;
the largest sample gap is1.002seconds. Filesystem device253:2 matches the earlier
flush-worker device spelling. At observer close the attempt has5,064,709,714 logical
bytes and5,065,228,288 allocated bytes (87 regular files).

| Host-wide observation | Delta or sampled value |
| --- | ---: |
| Device sectors written | 11,834,900 ×512bytes =6,059,468,800bytes |
| VM pages written | 1,480,424 ×4096bytes =6,063,816,704bytes |
| VM pages dirtied | 1,482,889 |
| kswapd pages scanned / reclaimed | 29,706 /28,826 |
| Direct-reclaim pages scanned / reclaimed | 0 /0 |
| Swap pages in /out | 87 /591 |
| Major page faults /OOM kills | 105 /0 |
| Minimum sampled MemAvailable | 14,519,376kB |
| Minimum sampled MemFree | 230,960kB |
| Maximum sampled Dirty /Writeback | 391,004kB /912kB |

Each arm dirtied about493–497thousand pages while the owning output footprint
grew by about1.7GB per arm. Reclaim/swap-out occurred in the third arm's interval;
neither was observed in the first two arm intervals. There was no direct reclaim
or OOM event in the complete observation interval. These are host-wide deltas,
not per-process attribution; available memory remained high and peak intervals
shorter than the sampling period can be missed.

The original controller is now unambiguously PID2354625/start204331337,
Python3.9, parent/group2354624, UIDs1000, CPUs0–7. PID2354654/start204331375 is
the retained pidstat observer. Pidstat shows the controller at8–9% in functional
pair windows, plus kernel housekeeping and the allowed live functional socket
observer. This explains how this new controller can appear as Python in pidstat;
it does not identify quiet5's expired PID2145712. There is no ownership exclusion
or retrospective grade change.

Persistence/writeback/reclaim pressure is a concrete lead beyond CDF file size.
A future isolated resource-budget experiment should preserve the full journal
workload and lossless originals while testing retention outside native windows.
This shared functional run cannot prove that persistence caused quiet5's tails,
and supplies no new timing comparison.