# Original QuickFIX/Go writer capability — 2026-10-19

Both native qualifications and independent cold reviews are complete. This
tests original-SDK serialization and parsing capability, with **zero timing
cells**. The full-bench plan remains incomplete.

## Scope and source

The latest retained microbench SDK pin is upstream
`12a9887734a755416a2b003dbe0ad48640dfa489`, selected read-only from its
clean source checkout. Original codeload archive SHA-256:
`aede310a4a1c77c396f2ddde48ec0bd2176117c88c438600c8631e8886e4a9d3`.
The 726-file content inventory SHA-256 is
`ba7059ea101398d0b233365fc0e536dd7d91324f59f97b02a023c0a50096340f`.
Every extracted SDK file is read-only; its bytes and mode remain unchanged
across module resolution, compilation and all actual executions on each host.
The shared SDK caches and existing adapters are untouched.

The public fixture uses `NewMessage`, `SetString`, `SetGroup`, `Bytes`,
`ParseMessage`, `GetString` and `GetGroup`. It constructs fresh messages under
both forward and reverse outer insertion, using SDK group templates. It never
supplies replacement FIX framing or body serialization. The native public
parser checks every scalar, group count and member value against its input.
An independent Python reader checks complete canonical values and group
associations, byte order, unique headers, BodyLength and checksum against the
shared specification, without accepting the producer's declared success.

Published producer/worker source: `7793165d637a276a325f4138f94429e7ed3dce56`.
Fixture SHA-256:
`70e6f5dfc1a7b706b537eb3ff4fb511206a3a3c4f9bc4085782e67c72be992e7`.
Fixed worker SHA-256:
`520eec7afd30050d7b91b6eb515517e6ff763570b7e198786dbde5c3686b8608`.
The independently generated input SHA-256 is
`0d4e2cb3c5124e1d70ebb44883c36b23b7ce4b308d525a9031fbfdd3a127c36e`.
The public source scrub has zero findings.

## Native execution and decision

Both durable user services finish successfully, UID/GID 1000. HP uses a high
priority isolated lease on CPUs 8–17, alongside shared conformance work; SCAN
uses a normal shared lease on CPUs 0–9. These are functional jobs without a
quiet timing claim. Both use the observed system Go 1.26.7 compiler,
`GOTOOLCHAIN=local`, `CGO_ENABLED=0`, `GOMAXPROCS=2`, `-p=2`, module verification,
`-mod=readonly` and `-trimpath`. All products are outside the repository.

Each host checks twelve sequences across all four flows: 120 canonical cases
and 240 actual serialized messages. Results are identical on both hosts:

| Message | Frames | Required values/groups | Canonical order | Order mismatch |
|---|---:|---|---:|---:|
| `D` | 72 | pass | 0 | 72 |
| `8` | 48 | pass | 0 | 48 |
| `V` | 48 | pass | 0 | 48 |
| `W` | 72 | pass | 0 | 72 |
| Total | 240 | pass | 0 | 240 |

For example, `D` requires body tags `11,55,54,38,40,44`; the public writer
emits `11,38,40,44,54,55`. `W` requires `262,55,268`; it emits `55,262,268`,
while preserving each group and its member order. In the pinned SDK,
`FieldMap.init` selects the unexported numeric comparator; the field-map
comparator, initialization override and production message build functions
are unexported. No supported custom outer-order API was found in this pin.

This is a **failed canonical writer capability decision** for the tested
configuration, with passing original-value/parser checks. A supported SDK
writer configuration or a separately reviewed new upstream source is needed
before fresh strict session/timing/physical qualification. The required Go
roster entry remains open; it is not silently excluded or relabeled a pass.

## Evidence and refusal checks

Both builds produce executable SHA-256
`9d6170fd4d4edacbc461b6820e7b09311b47ff1fd36d09cb1697965fdefe87c7`.
Native receipts:

| Host | Receipt SHA-256 |
|---|---|
| HP | `dc423779968039ee56dae77104796b91e286c1c941d8aa7432a94ddc47b1af58` |
| SCAN | `d1c6913f6a8f6baf2d48e7b5644458941fa4fd73aef643309323890e67a85410` |

Each executable rejects twelve malformed/incomplete inputs. Each cold reader
rejects nineteen alterations of copies of its real metadata, including
changed grade, owner/CPU/source identity, locks, source inventory/modes,
missing/refused native controls, duplicate metadata and symlink input. These
controls rehash changed envelopes to exercise the semantic checks. The
focused reference/refusal suite passes 34 tests, zero skips. A correctly
ordered independent writer passes the body checker; the current SDK's order
failure is an observation, not an expected-success rule embedded in tests.

Both services are `codex-fullbench-quickfix-go-writer-20261019-v1.service`.
Original host evidence is under
`~/build/codex-fullbench/libhft/quickfix-go-writer-qualification-20261019-v1`;
the bounded independently reviewed copies are under
`~/build/codex-fullbench/libhft/quickfix-go-writer-review-20261019-v1/{hp,scan}`
and retained in `~/archive/libhft/fullbench-20261019/quickfix-go-writer-v1`.
They contain functional wire vectors and provenance, with no measured latency
samples or CDFs. A first staging command failed Python parsing before either
service was created; the corrected, locally parsed launcher ran the unchanged
sealed worker. No root helper or sudo rule changed. No physical benchmark,
session pass, timing grade, comparison or report cell is granted.