# Canonical Java and native physical DAC matrices

2026-10-11. The user completed the home-directory installer. Both hosts now
hold the four requested immutable generations. Fresh physical matrices and
original failures are retained. **Every new row remains `S5 diag`; no
comparison, quiet timing or release gate is promoted.**

## Installed inputs and source identity

The three fixed installers recheck staged inputs and existing root destinations
on HP and SCAN. A separate read-only snapshot verifies 43 installed files on
each host, including launchers and counters: every descendant is root-owned,
is not user-writable and is not a symlink. Each host's three original installer
receipts are retained verbatim beside its independent snapshot and rechecks.

| Runtime | Installed generation / adapter SHA-256 |
|---|---|
| QuickFIX C++ | `fullbench-quickfix-cpp-canonical-v2`; `f725e4ae4c8ba5325d53ae63367e268017504307dcbe1200ea0ef91d97e2c507` |
| NexusFIX | unchanged `fullbench-nexusfix-canonical-v1`; `68cc831af7f0a9790ffa2783e714c8c9b7591db83b306a6b8a538c0fe71398cc` |
| QuickFIX/J | `fullbench-quickfixj-session-v4` |
| Artio | `fullbench-artio-session-v9` |

NexusFIX retains its original `ce5ae1074a25242cb4197f8e063105e01569d4f0`
compiled source and unmodified `ffa8b0a014b5cb4020ec8a99fe22fa3a24bfc494` SDK.
It excludes newer framework headers. These results do not qualify the separate
current-framework product or relabel the installed runtime.

| Complete controller capsule | Private source revision | Selected files |
|---|---|---|
| Java direct transport | `c0cea241b4efa7105f145dd01cc00a5719652f53` | 782 |
| Native direct transport | `26f3aea11415baa419b9088596ca96c663219553` | 783 |
| Counter retry | `7df504385019c7462d4d9d554ed15bca2cd6aa7b` | 786 |

Each contains the whole selected harness and three templates, with no engine,
internal source or Git metadata. Private deployment inputs have separate pin
and hash receipts. Every selected byte and mode agrees before and after the
runs. The controller prints `source-archive-unqualified`; its external manifest
binds the actual private source. All three qualified branches are pushed.

## Corrections, checks and failed preflights

Java and native selected-host commands now use Bash and local rsync. Other-host
commands retain SSH transport and original failure statuses. Actual command,
quoting, binary-copy/mode and failure controls pass. The final Java suite passes
**201 checks**, native transport **144**, and combined counter/controller suites
**353**, all with zero skips. Post-rebase logs are retained separately.

HP's first Artio matrix completed four tiny pairs with passing payload, admin
and raw checks, but immediate mixed/tick TX snapshots showed zero on both ports.
Its report was correctly refused. The initial native tiny NexusFIX tick showed
zero on one port; that failure remains retained. Its orchestration continued
after checking body/raw evidence without a positive tiny TX requirement. That
continuation does not admit the failed physical tiny pair.

Bounded read-only polling now saves every actual after-run counter snapshot.
It sends no probes, retains zero/partial reads and fails if both ports do not
advance within three seconds. Missing ports, changed interface/counter identity,
resets, malformed counts and boolean counts refuse the check. The fresh strict
HP retry requires every tiny physical gate before running its report.

Artio's retry tiny flows needed one, five, four and three reads respectively.
MD needed about **0.61 seconds** for both ports to advance; mixed and tick needed
about 0.47 and 0.33 seconds. All original zero reads remain beside final positive
snapshots. All four Artio and eight native retry tiny pairs pass physical,
body, admin and raw checks. Earlier failures are not regraded or overwritten.

## Physical results and host admission

Topologies are `hp/sfc-DAC/same-host-netns` and
`scan/mlx5-DAC/same-host-netns`. Each pair crosses one host's physical DAC;
no pair is labelled cross-machine. HP outer foreground jobs use high-priority
`isol-no-shared` leases and `wait` on busy exit 75; SCAN uses normal priority.
Slot status, processes, memory and IRQ-service state are retained. HP irqbalance
was active at these producer snapshots; this session changed no host service.

An unleased Java build/test started during SCAN's native lease; HP also had an
unleased Rust build at retry admission. The helper cannot exclude unleased work.
This record retains original numerical gates and leases but does not certify
actual quiet-machine timing or fairness.

Each completed report retains **30,000 primary and 30,000 scheduler-delay samples
per flow**, from three independent 10,000-sample process pairs at 10,000/s.
Warmups are 50,000 for Artio and 1,000 for QuickFIX/J/native reports. No raw
sample is dropped/clamped; missing offered-rate observations stay missing.

| Engine / generation | Host | Current numerical gates from original reports |
|---|---|---|
| QuickFIX/J v4 | Both | All four flows miss rate and delay; HP NOS also misses stability. |
| Artio v9 initial complete report | SCAN | NOS/tick pass local numerical gates; MD misses rate; mixed misses rate and delay. |
| Artio v9 counter retry | HP | NOS/MD/tick pass local numerical gates; mixed reaches 9,739–9,787/s and misses 98% rate. |
| QuickFIX C++ v2 initial/retry | Both | All four flows miss rate; delay failures vary by flow. |
| NexusFIX unchanged v1 initial | Both | All four flows pass local rate/delay/full-stability calculations. |
| NexusFIX unchanged v1 counter retry | HP | NOS/MD/tick pass; mixed retains a full-stability warning. |

The complete row index retains rate ranges and every explicit report blocker.
Some Java legacy process-status fields say `pass` below 98%; those are not the
current gate. All required body values, measured owner fold/count, admin and raw
series pass functional checks. Native QuickFIX C++ also proves both original
Logons. Comparison alignment remains declared false: `fixture_pass=false`,
and native aggregate contract status stays `not-audited`. Local numerical passes
alone cannot admit a comparison or `S5 pass`.

## Older queued runs and cold-check failures

Root markers woke older queued controllers. HP's older Java controllers failed
before pairs; independent exact probes reproduce self-SSH host-key failures.
SCAN's old QuickFIX/J completed four tiny pairs, then its partial archive's
unconditional Git lookup failed. Old SCAN Artio completed sixteen pairs.
Original artifacts, partial controllers, return statuses and later probes remain
separate. Original queued stderr was not recovered: later cause probes are
reproductions, not the missing original stderr.

The cold check first queued on SCAN shared CPUs. Only this session's waiting job
was cancelled before checking, then work moved to free isolated CPUs with shared
work allowed. The first actual cold attempt selected PATH Python 3.13 instead
of HP producer `/usr/bin/python3` 3.9.25. It refused an exact CV difference of
about 3e-18; raw observations, grades and blockers matched. A retry using 3.9
passed that comparison but refused the analogous SCAN comparison: SCAN's
retained controller argv declares `/home/yann/.local/bin/python3` (3.13.14).
Both failed attempts remain retained. The continuation selects each producer's
declared interpreter, records both runtime versions and preserves exact
comparisons; no tolerance or grade is relaxed.

## Cold proof and retained files

The independent continuation passes **164 original process pairs**, including
100 produced on HP and 64 on SCAN. It checks native original body/owner/admin
and Logon receipts, primary/delay raw hashes and statistics, every report
blocker, full percentile stability and measured load aggregates. Java checks
reproduce production payload/owner/admin receipts, process statuses and complete
primary/delay counts; original wire failures remain literal. No new pair or
timing observation is created.

It retains **720,480 native primary and 720,480 native delay samples**, plus
**480,400 Java primary and 480,400 Java delay samples**. All 6,070 inputs in its
inventory have identical hash, size and mode before and after the check.
Per-producer interpreters reproduce original numerical results exactly.
The twenty older automatically awakened pairs are retained separately; they
are not included in this current-controller cold proof.

[Committed evidence](evidence/canonical-physical-dac-20261011/) contains captures,
raw samples, original logs/argv/status, counters, source/runtime inventories,
failures and independent checks. Complete source tarballs and incidental old
controller Python bytecode/Git HEAD metadata remain archive-only, enumerated by
the import manifest.
The website archive consumes committed artifacts without changing their grades.

The durable archive seals **6,108 named inputs** (2,263,375,020 original bytes).
The final original-byte import contains **6,087 files** including seal/import
metadata; 23 source tarball/bytecode/Git HEAD files stay archive-only. The first
index check refused two older controller `.git/HEAD` copies, which Git cannot
represent as regular tracked files. Their original bytes remain sealed in the
archive; the import continuation explicitly classifies them archive-only and
verifies every other imported file is indexed, including 712 ordinarily ignored
raw/log inputs. No original archive input is changed. Seal SHA-256 is
`5161d211b1d06f9c9b852ea063ffbe248b8edf0169014bfff7a9013d4b3ed8d0`.
**Both HP and SCAN pass all 6,108 hash/size/mode checks**, with the same seal
and current import-manifest hash. The separate replication receipt and two
foreground verification transcripts bring committed evidence to **6,090 files**.
HP uses a high-priority shared lease; SCAN uses isolated CPUs with shared work
allowed. Both wait behind other owners when busy. These checks create no
benchmark pair or timing claim.

## Remaining gates

Full comparison-fixture source/runtime binding, rate/delay failures, the latest
NexusFIX mixed stability warning and actual quiet host admission remain open.
Current-framework physical timing, matching floors, other roster runtimes,
whole exported-consumer build/load qualification and release review also remain
open. **The full plan is not complete.**