# HP llfix with current framework headers, 2026-10-13

This generation checks the existing original-workload llfix TX wrapper against
current HftBench headers. The retained
[quiet5 timing generation](RECORD_2026-10-12_hp-framework-tx-quiet5.md)
uses an older framework ELF. Its raw observations, activity rejection and clock
limitations remain unchanged. This new generation is functional qualification
on ordinary localhost TCP, with shared host work allowed.

## Frozen inputs and method

Private source `d598c4336b8fdcfe2bb5c045c1378a8fa076820d` supplies sixteen
unchanged files: two original-workload wrappers, three Python controller/auditor
files, two Python test files containing seven checks, all eight HftBench headers,
and the assertion-enabled native sample-storage test.

Compared with the original framework preparation, only `HftBenchContext.h`
and `HftBenchRunner.h` differ among the eight headers. They reserve sample
storage before measurement and expose cold task preparation. The clock,
latency statistics and raw sample format remain byte-identical. The wrapper
is not edited, and its primary probe remains local construction/send, not
round trip. No performance effect is inferred from the source difference.

The existing root manifest
`78abcf4109fa5c0b29c96feeb5656bd507caae15fa6a3e842135b7e36c9b4b17`
binds the original 204-file llfix 1.0.7 source tree, GCC 11.5 and receiving peer.
The SDK remains at upstream `165cc999fa283a4ead17a7b935c5d3f850fec851`.
Source is read from the retained archive tree; removed disposable preparation
directories are not recreated or silently treated as available.

A fresh out-of-tree TX build uses the same C++17/O3/native flags. A second
build runs the existing seven-case allocation test with assertions enabled.
Both dependency files record actual compiler inputs. Preparation and cold
review verify those inputs, all frozen sources, the original SDK and peer.

High-priority shared leases cover staging/build and cold retention. A separate
high-priority isolated lease permits shared work for tiny functional pairs;
it makes no quiet timing claim. Native client/server use CPUs 17/16, with
controllers on CPUs 0–7. The existing ready barrier allows independent
PID/start-time/executable and every-thread affinity checks before release.
The irqbalance service must read loaded/inactive/dead/PID0 before and after
native execution; it is not changed.

The planned four pairs cover both original RDTSCP/CPU-MHz and MONOTONIC_RAW
clocks, with measured/warmup counts 128/1 and 257/17. Cold review requires
every persisted frame's envelope, required numeric values, groups and
contiguous owners; exact bilateral application bytes; 806 orders per peer;
770 measured samples; and full-width CSV/HFTBRW count/statistic agreement.

## Outcome

The [preparation-only input seal](evidence/hp-llfix-current-framework-preparation-20261013/selected-files.json)
retains all sixteen frozen source files and exact staging, worker, build,
functional, cold, retention, replica and verification operations. Their Python
syntax is checked. All seven existing Python checks pass locally, with zero
skips; their actual output is retained separately from the pending HP results.

The frozen source is backed up on private ref
`s/codex-llfix-reference-hp-current-framework-source-20261013`.
At 22:56 UTC, HP staging request `67399362644f46179e067291b67c0849`
was queued shared/high behind another owner's preflight and the already
queued full-session physical report. No owner process was changed or bypassed.
That preparation checkpoint claimed no native pass or performance result.

## Staging continuation, 2026-10-14

The next read finds an incomplete transfer: six original source files, totaling
56,235 bytes, exist under `hp-current-framework-20261013-ops30`. The manifest,
worker script, launch receipt and products do not exist. No build or native
pair started. The old waiter is no longer active and there is no reliable
original exit receipt; the evidence does not attribute the interruption to a
particular process or timeout. The six files and original queue log are kept.

Private source `f35fe936e492b912696602e84fad2293b5491c4e` adds a small capsule
validator plus six regression checks. A bounded compressed argument replaces
the live SSH input stream. Every member's path, type, mode, size, digest and
complete roster must validate before the output root is created. Truncation,
changed bytes, missing/extra/duplicate members, unsafe paths and symlinks refuse.
All six local checks pass with zero skips. The owning helper neither requests
leases nor starts benchmarks; it changes no shared helper API.

The [staging evidence index](evidence/hp-llfix-current-framework-staging-20261014/selected-files.json)
retains the exact new operations, packet manifest, original failure inventory,
local checks and real remote process observation. The complete local compressed
capsule is 46,503 bytes, carrying 23 files in a 62,004-character argument. SHA256:
`e09deaf858a61f9425f31454f1a6b3301d1f0b7b60d093b03dce2adab8225993`.
The sixteen framework inputs remain byte-identical to the original frozen pin;
the additional source files are only the staging helper and its tests.

At 01:37:12 UTC, HP starts detached waiter PID 3300275/start ticks 218692040.
At 01:39:37 UTC, a separate SSH read confirms parent PID 1, stdin `/dev/null`,
and the complete compressed packet and 23-member manifest in its arguments.
Their actual digest equals the local capsule. The source packet no longer
depends on keeping the launching SSH process connected. Source extraction,
builds, native pairs and retention still require their declared CPU leases.

Request `0f5b0757ca974e1abe0e9cf773f6e159` waited shared/high behind existing
work. At that checkpoint, preparation and native qualification were pending.

## Preserved preparation failure

At 02:39:46 UTC the complete 23-file capsule staged and launched its worker.
Preparation then failed before any build: the operation still requested
`manifest30.json`, while its actual input was `manifest32.json`. This was an
operation-script filename error. The worker skipped native execution and
retained the original failure successfully at 02:39:47 UTC.

The [failed-generation evidence](evidence/hp-llfix-current-framework-20261014/selected-files.json)
preserves the unchanged failing script, traceback, complete source/manifest
and terminal worker receipt. Its complete 34-file compressed archive is
81,550 bytes; both HP and the local streaming verifier reproduce its bytes,
sizes and modes. Archive SHA256
`6189cd8870312d6204873718acb2367d4dbcd9578ddd36c7ded3129493d3499f`;
seal SHA256 `c7994db0477ca660033d4e5defa91139cdc55629d70e297132155abbaeec281b`.
No failed receipt or original source is overwritten.

## Fresh build and functional qualification

The continuation uses fresh root `hp-current-framework-20261014-ops33`,
retaining the internal manifest/source/receipt filenames. It corrects the one
stale manifest lookup and selects a distinct archive name. Before queueing,
a local check stages the complete actual capsule and executes the frozen
`inputs()` function body: the original missing-file failure reproduces, the
corrected body reads all eighteen exact sources, and a changed-source control
refuses. This runs no compiler or native process. Its original control receipt
and exact operation sources are retained with the successful continuation.
All fourteen operation functions match after normalizing the four declared
filename/root/archive strings; twelve function ASTs are directly identical.

HP launches the new detached worker at 06:19:55 UTC. Preparation finishes at
06:20:06 UTC; four actual unprivileged localhost pairs finish at 06:20:09 UTC.
Independent cold review passes at 06:20:25 UTC. All phases use high-priority
CPU leases; the functional isolated lease allows shared work. Controllers and
auditors stay on CPUs 0–7. Native clients/peers use CPUs 17/16; each process's
PID/start ticks, executable bytes and every thread's actual affinity pass the
ready-barrier check. The irqbalance service reads loaded/inactive/dead/PID0
before and after; no service or other host policy changes.

Both actual GCC 11.5 builds pass, and the dependency capture binds 522 compiler
input files. The unmodified 204-file llfix 1.0.7 source tree and original peer
remain unchanged before/after. All thirteen Python checks pass without skips.
The assertion-enabled native storage test passes seven cases, including zero
sample-storage allocations during its synthetic measured-loop scopes. This
does not claim that the llfix SDK itself performs no allocations.

Fresh TX ELF SHA256:
`e94546c82dd0d0792d129c6ac2aa63b05ad6437454ca8d413ac1dd7db49686bd`.
Native storage-test SHA256:
`876b8cdce6695e804a61bfee44f5cbac16a67c1ffc25b774dc83792b22e50aaa`.
The receiving peer remains
`9f7a2379b99da6c7df7216f276bde074c30b72de0d27fcae902cb278677f1cf4`.
Only Context/Runner differ from the earlier framework header map; all sixteen
original wrapper/controller/header/test source inputs retain their frozen bytes.

| Clock | Measured orders | Discarded warmup | Matching orders per peer | Cold result |
| --- | ---: | ---: | ---: | --- |
| Upstream RDTSCP/CPU-MHz | 128 | 1 | 129 | pass |
| MONOTONIC_RAW | 128 | 1 | 129 | pass |
| Upstream RDTSCP/CPU-MHz | 257 | 17 | 274 | pass |
| MONOTONIC_RAW | 257 | 17 | 274 | pass |

All **806 original application frames per peer** pass strict envelopes,
required converted values, both party groups and contiguous owners. Their
complete ordered serialized bytes match bilaterally for each pair. All **770
measured samples** pass full-width CSV and HFTBRW/count/statistic checks,
including individual extrema and the warmup exclusion. Neither raw data nor
historical performance values are rescaled or downsampled.

This qualifies the new build's ordinary-loopback functionality and sample
retention. It supplies no quiet timing, Onload/physical, vendor reproduction,
Stage 2, SDK release or current-framework performance grade. Installed root
payloads remain unchanged; historical timing still refers to its original ELF.

## Complete continuation retention

HP seals all 121 continuation files at 06:21:46 UTC, including all raw journals,
CSV/HFTBRW samples, source, binaries and receipts: 5,907,367,641 logical bytes in
a 26,312,268-byte lossless compressed archive. Archive SHA256
`8bd7b9bc6c541a498de01fb32fef50d6c9984b6719ae29ab935638ff3858b7d6`;
seal SHA256 `9ef2f1b801c6899d0c513e03964fb421d8cc6adcbf2a1b2c4735e520475217b8`.
HP and the local streaming verifier reproduce every member's bytes, size and
mode. The [qualified compact evidence index](evidence/hp-llfix-current-framework-qualification-20261014/selected-files.json)
keeps the complete source, build, wire/raw, identity and outcome receipts; all
35 omitted raw/binary files remain in both complete lossless archives. No raw
journal, sample series or binary is added to Git.
Both hosts retain the archives under
`~/archive/libhft/codex-llfix-reference-20261014/`: `hp-current-framework2.tar.gz`
is the original failed preparation, and `hp-current-framework3.tar.gz` is the
successful continuation. Both compiler dependency files are included in the
selected evidence.

The first local materialization refused an existing generic `worker32.log`
from the earlier failed generation. The received archive was complete and
unchanged. Its failing source, executor traceback and receipt are preserved;
a separate local continuation reads the same container, verifies the already
received archive bytes, and writes the new worker metadata beneath its owning
generation directory. It starts no new native pair or socket. The original
26,327,040-byte received container remains losslessly compressed in 699,268
bytes, with its original digest/size/mode and full compression proof retained.

The compact cold verifier reproduces the four actual successful pairs,
eighteen frozen sources, thirteen Python checks, seven native storage cases,
all identity/affinity checks and original frame/sample counts. No original
failure is regraded. Additional source/control receipts are labelled outside
the immutable native evidence seal.

A guarded high-priority shared cleanup waits behind another owner's Rust
suite, starts at 06:40:23 UTC and is complete by the next lease at 06:43:08 UTC. Its
[completed receipt](evidence/hp-llfix-current-framework-qualification-20261014/postcleanup/staging-cleanup32.json)
verifies both complete compressed copies and each selected source/staging byte,
then removes exactly sixteen large staging duplicates: **5,905,580,032 logical
and allocated bytes**. All active originals, both full archives and the seal
remain unchanged. A separate read confirms every removed staging path is absent
and every active source file still exists at its original size.

Only this session's waiting cleanup ticket was replaced with a detached waiter,
before any cleanup started. The replacement uses complete arguments and stdin
`/dev/null`; no other owner's job, queue or lease was changed. Exact launch,
helper, completed cleanup and administrative-cancellation receipts are retained
outside the original native evidence seal.

No Stage 2 adapter, shared framework/clock/helper, SDK, root runtime,
namespace launcher, service, IRQ, turbo or other host policy is changed.
Earlier vendor and timing grades remain unchanged.