# NexusFIX original-byte comparison fixture admission

2026-10-12. A **new source generation** qualifies its cold and production
comparison fixture on both HP and SCAN. This is localhost functional evidence.
**No new physical run, report timing, installed-runtime or release admission.**

## Implementation

The new driver declares `canonical_body_v2` at both endpoints. Its extraction
model remains the existing `canonical_spec_v1`; payload values, public SDK and
shared clock APIs are unchanged. The native controller grants fixture admission
only after independently checking every original warmup/measured body,
complete bilateral byte identity, FIX framing/checksums/increasing sender
sequences, the common owner sink/fold count, prepared tick boundaries and
measured administration with original thirty-second Logons. Both endpoint
fixture declarations must agree. A runtime's printed declaration alone cannot
satisfy this gate.

The cold reader additionally requires all 120 original SDK frames to match,
all 48 cycles and the 57-control OrdType rejection set. A new declaration with
normalized prices, extra fields, different field order or incomplete controls
fails even in the default reader. Exact legacy diagnostic vectors remain
unadmitted. Cold/production fixture admission supplies no sample-rate,
scheduler-delay, stability, physical-link or quiet-host admission.

Controller JSON and CSV retain `fixture_observation_status` and the declared
comparison generation. Every outer run must pass; a good new run cannot lend
admission to a legacy or failed run. Installed NexusFIX v1, all older runtime
pins and original literal false flags remain unchanged. No historical row is
regraded as a fixture or Stage 5 pass.

## Source and real checks

| Input | Identity |
|---|---|
| First new admission source | `57c2f76e66e05be048b23f3c6c6ba3c5880edaa5` |
| Corrected complete archive source | `a8b5a6b7b59b7cccd3fc34e41a1fddf348d02df6` |
| Unpatched external NexusFIX SDK | `ffa8b0a014b5cb4020ec8a99fe22fa3a24bfc494`, 91 pinned headers |
| Fresh production adapter on both hosts SHA-256 | `2ebc91a199aec2c7c796dda66a71e743f7dc538626b7966b538c00f26ee0ccee` |
| Fresh cold fixture on both hosts SHA-256 | `5a630347e407425a5eccbd28a9028ab14c7ad7d3001438ddfd85bf46e50732cf` |

**185 focused checks pass with zero skips**, in development and on each host's
complete 789-file stripped archive. The archive contains no engine/internal
source or Git metadata and is extracted inside an unrelated Git repository;
the controller retains `source-archive-unqualified` rather than borrowing that
repository's identity. The SDK build receipt binds all fifteen actual current
adapter/fixture/framework inputs and the original compiler argv/version.
Source bytes/modes and the external SDK remain unchanged after qualification.

The first archive exposed an existing ordinary-remote-mode dependency on
private DAC namespace facts: **183 checks passed, two failed**, before any
benchmark pair or adapter compile. Development's private config had hidden the
problem. The controller now loads namespace facts only in DAC mode. Both
remote-copy regression cases explicitly run with no private rigs, including
failed retrieval with stale artifacts. The first failure and exact first
source archive are retained. Only the verified obsolete own HP wait PID
1993453 was cancelled; no other job or queue was stopped.

Host CPU-slot status, running processes and memory were checked before work.
SCAN's outer foreground build/tests/functional suite waited in the `isol`
pool, with shared work allowed; this grants no quiet timing. HP's corrected
suite waits in `shared` with **high priority**, then completes after another
owner's exclusive job. Original preflight/queue snapshots are retained.

Each host freshly compiles the SDK-bound production adapter and cold fixture.
The actual portable controller completes **eight independent localhost TCP
pairs per host**: all four flows with warmup zero/one inner run, then warmup five/two
inner runs, twelve measured iterations per run. The second CLI verifies reuse
of that exact binary and its SDK receipt. Both original CLI outputs remain
separate. All sixteen controller fixtures pass; CSV and JSON rows match exactly,
retain actual achieved/completion/deficit observations and leave unmeasured
offered rate empty. Each row remains `S5 diag` because it is a tiny local run.

Each host retains **144 primary and 144 scheduler-delay samples**, without drops
or clamping, and **460 original application frames per peer**. All body,
owner and admin proofs pass. The cold fixture separately passes **48 cycles,
120 original frames and 57 malformed-payload controls**. An independent
read-only continuation from the same complete archive exactly reproduces all
sixteen production observation receipts and the cold fixture proof, preserves
every original input byte/mode, and creates **zero new socket pairs**.

## Retention and remaining work

Originals, source archives, failed checks and compiled products are sealed in
`~/archive/libhft/nexusfix-fixture-admission-scan-20261012` and replicated to HP
and SCAN with exact SHA-256, size and mode checks. The import excludes only
compiled products and complete source tarballs from website downloads; these
remain in the archive. Import and replication receipts describe every file.
The first queued HP log is an explicit snapshot. Its later completed results
are retained in a separate HP supplement; the first SCAN seal is immutable.

The first SCAN archive seals **137 files / 6,675,012 bytes**, with seal SHA-256
`5590610d5f04d037edf2310806eb4c4d77f874f366448abdd50f2fb44f04942b`.
Its import contains 135 files plus a replication receipt; four source tarballs
or compiled products stay archive-only. HP's completed originals, joint cold
proof and new staged payload form a separate immutable supplement at
`~/archive/libhft/nexusfix-fixture-admission-hp-followup-20261012`, with its own
import, retention seal and bilateral replication receipt. The supplement seals
**115 files / 1,679,531 bytes**, with seal SHA-256
`777122509606730a0cac3f9f84ec96f3336fba4b178cc6f8ba3902290aab3b1d`.
Its import contains 114 files plus a replication receipt; three compiled
products remain archive-only.

The separately pinned `fullbench-nexusfix-canonical-v2` payload contains five
files: the identical binary independently built on both hosts, fixed launcher,
SDK build receipt, bilateral fixture admission and provenance. Both staged
payload checks pass. The private machine config retains the existing v1
observation schema capability and every unchanged QuickFIX C++ v2 pin. Its
new directory coexists with NexusFIX v1; an existing mismatched v2 refuses
installation. The fixed root installer checks all hashes, copies to a temporary
directory, rechecks root ownership/modes and atomically installs it.

Seven root-runtime installations, including this generation, now await user
sudo authentication through `~/install-fullbench-stage5-next.sh`, followed by
fresh DAC evidence. Existing installed
v1 and historical reports stay immutable. Full physical rate/delay/stability,
quiet-host parity, whole-harness standalone and release gates remain open.