# NexusFIX retained scheduler delay and maximum sample reporting

The full-session adapter now retains the shared harness's `scheduler_delay`
series alongside its primary raw series in paced runs. Rows declare the delay
sample count, clock API and all measured runs scope, with the mean of per-run
p50/p99/p99.9 values. Closed-loop controls have no delay series and report zero
delay samples. Both primary and delay `worst_us` values use the actual maximum
retained sample across measured runs. The former primary row used a mean of
run maxima. Raw output still uses the shared HFTBRW1 writer and concatenates
measured runs; neither the writer API nor pacing clock/workload changes.

Failure to create a requested raw output now raises an adapter exception rather
than returning a successful row without its artifact. The driver's established
exception exit code is **99**. Production application capture and prepared
tick clock boundaries remain on the already qualified public SDK paths.

## Fresh HP qualification and separate continuation

Tested source **`2b0eb49e1ec2542d40192cc4303d8f5f2fc4a145`** is retained on
`s/codex-fullbench-nexusfix-delay-source-20261003`. A fresh immutable capsule
contains **722 source files** and **91 unchanged upstream headers** at SDK pin
`ffa8b0a014b5cb4020ec8a99fe22fa3a24bfc494`. HP's status/process/memory check
precedes a high-priority `isol` CPU lease allowing concurrent shared work,
with the outer foreground controller on CPUs 8–17. Clang 21.1.8 builds the
driver and fixture with C++23/O2, DNDEBUG and AVX2, without SDK patches.
Both builds, the adapter self-test, **48 cold cycles**, **51 malformed-payload
rejections** and **32 actual TCP pairs** pass, with zero skips.

The pairs retain the previous twenty closed-loop cases and add twelve paced
cases: every flow at `(warmup,runs,samples,rate)` values `(0,1,12,1000)`,
`(5,2,12,2000)` and `(3,3,20,1000)`. All owners match the independent common
sink/fold model. All **1,264 primary samples** and **384 scheduler-delay samples**
are retained. Delay counts exclude warmup; independently recomputed nearest
rank percentiles and cross-run means match the printed values to rounding.
Primary/delay printed maxima match the maximum raw sample. Twenty-eight
captured pairs retain **3,370 byte-identical application frames per peer**,
with production envelopes/session sequences and required values checked.
Four cases remain capture-disabled controls.

The original outer checker stopped at its first negative raw-output control:
the client correctly returned 99 while the checker expected 1. Its original
failed report, runner and logs remain unchanged. A separate continuation
regrades all 32 completed pair receipts, verifies the actual exception and
absent raw file, then runs the remaining server-owned tick negative control.
Both missing-output controls reject with 99 while the peer exits successfully.
The continuation passes and verifies that every original file, all source
bytes and all SDK headers remain unchanged. No successful pair or build is
repeated; the original outer receipt retains its failed grade.

Driver SHA256:
`252899f64cf4954f4156e587e6a200d662ffe67c8f731f6811dd184bd80d7b5d`.
Fixture SHA256:
`fc31c761874cd54318a7c0454ec3abdc6b554ea3fc441523b1114138769cbc94`.
Compiled products and the original/continuation receipts remain under
`~/archive/libhft/nexusfix-delay-hp-20261003-2b0eb49e1ec2` on HP.
The source/scenario/SDK checks pass **82 checks**, with zero skips; the source
comprehensiveness gate passes.

[Sealed original checker/build/functional receipts, continuation, SDK/source
inventory, bilateral wire captures and raw samples](evidence/nexusfix-delay-20261003/hp-2b0eb49e1ec2/SHA256SUMS).

## Remaining gates

These are tiny localhost functional diagnostics at 1k/2k offers per second;
they establish delay retention and accurate extrema, without timing, report
rate or physical link admission. Full SDK body alignment and the common
OrdType/dictionary choice, complete admin observations, physical DAC rate and
stability, and release qualification remain open. Every row retains
`fixture_pass=false`; no installed payload pin or historical result changes.