# HP archived framework idle cold review

2026-10-11. This is an offline review of the already completed
[idle-retry generation](RECORD_2026-10-11_hp-framework-tx-idle-retry.md), using HP
shared/high. It starts no native pairs, writes no service or host policy, and
does not visit SCAN. Turbo remains deferred. Fullbench's current Java physical
matrix queues behind the existing shared audit through the host CPU queue.

## Exact old inputs

The original source remains `0a7c09c3f75a9b4696f1bef309c5ad0746b1024c`, with
controller SHA256 `7993e4b22cfb68070920f01ce93e02186d033a477542c542591a1e8a13813600`.
The `functional2` capture remains SHA256
`679022d2ed67b05c3c3c718a8ef7ace5063f1968e73f3d634f4b07cc691b160e`.
Its status remains failed: requested IRQ167 changed 3 to 7 and IRQ172 changed 1 to 7,
while their effective masks stayed 3 and 1. No ordinary audit or timed matrix is
created. This generation is not an eligible positive functional proof.

All 104 sealed input files are verified and materialized under a fresh owning
scratch directory. The removed old source/runtime paths are not recreated.
Only the three pair directory strings in a disposable in-memory capture copy
are projected to the new materialization. Reversing those three substitutions
must reproduce the original capture object exactly. The original capture,
preparation, source maps, native identities, groups, counts, policies, socket
observations and sample bytes remain unchanged.

The five native/wire/raw statements are taken verbatim from the original
controller's audit and checked for exact AST identity. Its policy admission and
success promotion are not executed. The current installed source/compiler,
runtime, binary and protected-launcher bindings are independently rechecked;
this does not substitute a new controller or current framework binary.

## Cold review

All three original/framework-upstream/framework-monotonic arms pass the cold
native checks. Both peers' journals contain 1,001,000 successful orders per arm:
6,006,000 independently inspected order copies in total. Framing, BodyLength,
CheckSum, converted required values, both parties, owner sequence and counts
match. No failed order or execution-report reply is accepted. Saved native
placement, original socket identity/options/buffers and positive physical packet
deltas also match their capture.

Both million-sample framework arms pass exact uint64 CSV/HFTBRW agreement and
recomputed summary checks: 2,000,000 raw samples in total. The original summary
and saved divisor match its original log. Every one of the 104 materialized
files retains its original hash, size and mode after review; the compressed
original verifies again unchanged. The three in-memory directory projections
are recorded separately from the immutable original capture.

The diagnostic status is `functional-wire-raw-pass-policy-rejected`, with
`eligible_functional_proof=false`, `timing_admission=false` and
`vendor_reproduction=false`. Native checks SHA256:
`9eb0118d91d423afbdd2ad7919b58acd98fdb58a28fdc1af18928bd7b7b4e470`.
The three-divisor agreement does not override the policy rejection.

The first archive-review bootstrap used a helper path missing its
`onload-observer-source1` component and failed before materialization. Its exact
operation and error log are retained. The corrected operation verifies both
frozen helper hashes before reviewing data. No original failure is rewritten.

## Retention

After fullbench's existing Java/native exclusive matrices completed, the compact
proof was sealed on HP and independently verified locally. Archive
`~/archive/libhft/codex-llfix-reference-20261011/hp-framework-idle-cold1.tar.gz`
has 30 sealed files and 141,854 compressed bytes. SHA256:
`7a571d2e5dc281cef361d704dd5dde4a05553974431b4a3b184d93fb4f8e5247`;
seal SHA256:
`699918ec8812e610c782ffdb4b76cf60eab131a614f0cd131789a4c4fcb71948`.

The large journals and samples remain in the already verified 104-file original
archive, SHA256
`7c16a4733254037d49550ffc8f60affc100f08a82e60658ee456c90d7785aa88`.
The compact proof retains 14 exact original source/preparation/capture inputs,
all review and replica receipts, original native check statements, directory
projection and operation scripts, including the initial bootstrap refusal.

[Selected exact files](evidence/hp-framework-idle-cold-20261011/selected-files.json)
retain the complete compact archive and postseal verification metadata. The 104
file materialization was removed only after complete byte/mode checks, full
original-replica verification, original native identity exits and readable
same-UID consumer checks. Inaccessible process fields and limited inspection
scope remain explicit. Final cleanup then removes the 24-file compact scratch
and 31-file uncompressed archive duplicate after both replicas pass and their
complete membership/hashes/modes match. It checks readable same-UID consumers,
retains eight inaccessible process observations without claiming root inspection,
and confirms the active source3/ops4/root3 preparation/zero-native guard and home
helper remain intact. The 38 selected exact files total 542,191 bytes. Both
compressed archives remain on HP and locally. No own job remains queued.

## Separate current policy guard

The new `09906965356e1eab39b3a593a35c8d2ec50354ef` source generation and prepared
epoch remain intact. They require irqbalance loaded/inactive/dead with PID0
before native work. Current readback is active/running, PID1325. The earlier
zero-native guard refusal and actual password-required sudo refusal remain
retained. This old-data review does not qualify the new guard's positive arm
or permit its quiet capture.

The operator action remains, as `yann` on HP without a leading sudo:

```sh
~/pause-irqbalance-for-bench.sh
```

The helper requests its own exclusive high-priority host lease. Fullbench's
root-installed generations remain separately owned and pinned.