# Native full-session load summaries and current standalone build

2026-10-11. Summary retention, standalone build and functional qualification.
**No new physical timing, comparison-fixture or release admission.** Original
controller records and pending immutable root payload pins remain unchanged.

## Summary change

Canonical native JSON rows previously left achieved rate, completion and
deficit empty even when the owner emitted observations. CSV also discarded
those fields, offered rate, capacity values and the benchmark layer. The new
summary retains them across JSON, CSV and Markdown, with the configured paced
target distinct from any measured offered rate.

Every declared outer process pair contributes, including weak and failed
pairs. Rate/completion aggregates use the minimum; ranges and observed/expected
run counts disclose incomplete observations. Missing, nonfinite, malformed or
negative measurements stay missing, while a genuinely measured zero remains
zero. An aborted outer run set cannot produce a complete aggregate.

Completion uses a valid owner percentage or the owner's completed samples
divided by completed samples plus its observed deficit. Its source is retained.
Missing deficit or offered-rate counters cannot be inferred from the profile
or target. Closed-loop rows have no paced target. Existing original raw/admin,
fixture, physical and report gates are unchanged; a complete load observation
does not itself admit a comparison or Stage 5 row.

## Source and checks

| Input | Identity |
|---|---|
| Frozen source / complete stripped archive | `9ef50f8e290bcffffd926cc611986885ea4abe4e` |
| Preserved private branch | `s/codex-fullbench-native-summary-qualified-20261011` |
| External unpatched NexusFIX SDK | `ffa8b0a014b5cb4020ec8a99fe22fa3a24bfc494`, 91 pinned headers |
| Fresh production adapter SHA-256 | `fd34fc367d01e411c2c539c8fd15161405a4a1326fabbb3b7bacd0dfe3d90c51` |
| Fresh cold fixture SHA-256 | `cfc0b340da8f83c48dd3ba9ca3f69c2c35549bfdde096db321df5bcc392ead39` |

**368 focused checks pass with zero skips.** The new twenty checks exercise
production aggregation and CSV/JSON/Markdown serialization, the separate
canonical/TX/capacity/closed-loop labels, actual zero, incomplete observations,
partial completion, missing/invalid counts, aborted outer runs and retention
of all failed physical runs. Those physical fixtures are synthetic; they are
not physical evidence. Existing raw/admin/body, quiet/stability and explicit
fixture-blocker controls still refuse admission.

The complete selected archive has **781 files**, including three export
templates, with no engine/internal source or Git metadata. It is extracted
inside an unrelated real Git repository. The actual controller retains
`source-archive-unqualified`, without borrowing that repository's identity.
Readiness has zero scrub/reference/link findings and 66 textual dependency
escapes; whole-harness source/runtime and release qualification remain open.

## Original production observations

A read-only local continuation consumes the separately sealed
[original eight NexusFIX pairs](../../../private/benchmarks/retained/records/RECORD_2026-10-11_nexusfix_complete_controller.md).
It independently reproduces all body/owner/admin receipts, preserving the
original 144 primary/delay samples and 460 application frames per peer. Every
original file hash, size and mode agrees with that generation's existing seal
before and after the check.

Each original owner row passes through the new real summary/serialization
functions. The resulting CSV and JSON rows match exactly, retain the original
achieved rate and observed deficit, and correctly leave unmeasured offered
rate empty. All eight rows keep their diagnostic/fixture/report outcomes.
This creates no socket pair, new latency observation or new timing grade.
The stripped archive's twenty new checks also pass without skips.

## Fresh current-framework qualification on HP

The earlier reused runtime explicitly excluded two newer shared framework
headers. This generation freshly compiles the adapter from the current
stripped archive and binds **all fifteen actual adapter/fixture/framework
inputs** in the SDK build receipt. This includes current `HftBenchContext.h`
(`2036bd3d81f0f4c443f44b9eb4db570d29c760e26380af452b27cd3bf486852a`)
and `HftBenchRunner.h`
(`551d995627c99ac3a5cccb777af41ad96353564ca37d7ed60206c03826819311`).
It does not relabel or replace the older producer/runtime.

HP host slots, running processes and memory were checked before work. The
outer foreground build/tests/functional suite used `host-cpu-slot run --pool
isol --priority high`, with the same `wait` command on busy exit 75. It acquired
its slot directly. Shared work was permitted by that pool; no quiet timing,
disjoint endpoint placement, physical link or host-policy change is inferred.
Clang 21.1.8's original version output and exact compiler argv are retained.

The actual portable controller builds the new SDK-bound adapter and completes
**eight fresh local TCP pairs**: all four flows with warmup 0 / one inner run,
then warmup 5 / two inner runs, twelve measured iterations per run. The second
CLI uses the adapter's matching SDK receipt to prove exact binary reuse; the
first CLI's original files remain retained separately. Both actual archive
summary suites pass twenty checks, zero skips.

The actual pairs retain **144 primary and 144 scheduler-delay samples**, with
no drops or clamping, and **460 original application frames per peer**. Full
original body, owner fold/count, admin-window and prepared tick-boundary
receipts pass. Actual original owner rates survive into identical CSV and
JSON rows. All samples are completed, with observed deficit zero and no
invented offered rate. Their `required-values-diagnostic`, `fixture_pass=false`,
`S5 diag` and unrequested physical report gates remain literal.

A separately freshly compiled cold fixture passes **48 cycles and 57 rejection
checks**, requiring exact D/V/W/8 bodies and the limit-OrdType rejection set.
All 91 SDK headers and every selected source byte/mode remain unchanged.
This qualifies the current standalone build and functional path; it does not
qualify a current-framework timing report or every other exported consumer.

## Original checker and wrapper failures

The retained original failures remain separate from their corrected checks:

1. The first focused run passed 57 checks and failed five because the new
   mock used the remote call's run-index position for a local call. Corrected
   local/remote indexing passes all 62 focused checks. No production adapter
   change was required by this test-harness failure.
2. The first archive preparation missed the exporter helper import path; its
   dependent check refused the absent capsule. No archive, SDK build or pair
   was consumed. The corrected preparation retains the full frozen selection.
3. The first cold replay passed its twenty archive checks but used the wrong
   retained case path. It failed before any original pair was consumed. The
   continuation uses original `run_1` directories and separately retained
   owner logs, and reproduces all eight original receipts.
4. The first HP queue wrapper had a missing shell quote and exited before
   taking a lease or compiling/running an adapter. Its original script and
   transcript remain. The corrected syntax-checked wrapper completes the
   fresh build and eight actual pairs above.

## Retention and next gates

[Committed evidence](evidence/native-measured-summaries-20261011/) retains the
source selection, original checks/failures, exact fresh process logs, captures,
raw files, cold vectors, SDK build receipt, summaries and manifests.
`~/archive/libhft/native-measured-summaries-20261011` seals **157 named inputs**;
the complete source tarball and two compiled products stay archive-only.
Seal SHA-256 is
`eb1a7f50da4e6114d0d97fb6685af5fa5e5e3a3ebd1acf03f2393df6063d7393`.
Both HP and SCAN verify every named hash, size and mode and the seal itself;
the committed replication receipt retains those checks.

These new products are separate cache/archive inputs, not root-installed
payloads. Existing native-v2, QuickFIX/J-v4 and Artio-v9 staged installers
remain unchanged and still require the pending password-protected root
action on both hosts. Comparison-fixture source/runtime binding, fresh quiet
physical matrices and rate/delay/stability gates, whole standalone consumer
qualification and release review remain required. The website archive can
retain these diagnostics without promoting them.