# C++ CORE and NexusFIX matched physical review

The maintained SDK-free reader now independently replays both full arms on HP's
Solarflare DAC and SCAN's mlx5 DAC before matching them. These remain same-host
physical-link namespace topologies. Both profiles contain 1,000 warmup cycles,
10,000 measured cycles per run, three runs and a 10,000/s offered rate.

## Actual result

NOS/ER, market data and mixed flow pass matched admission on both rigs. Every
original body, typed owner/count, administration artifact, raw primary/delay
quantile and true extremum, deployment/source binding, quiet lease and endpoint
CPU mask is checked. Original physical counter identities and positive deltas
also pass. No SDK, socket or physical run is added by this reader.

The tick comparison is **blocked on both rigs**. CORE stamps receive before
parsing, while NexusFIX stamps application callback entry before extraction.
Those spans differ. Both original producer grades remain preserved, but the
comparison emits no tick ratios and cannot admit a four-flow C++ headline.
Aligning a newly qualified runtime's receive boundary is the next concrete gate.

The full reader covers 48 original full physical pairs across four arms:
480,000 primary samples, 480,000 delay samples and 1,320,000 original application
frames per peer. Across-run percentile comparisons use medians; worst uses the
maximum of every actual run extremum. Twenty-seven source/shape/counter/clock,
metric and relocation controls pass, zero skips.

## Portability corrections and retained attempts

The first reader stopped on the native schema's three joined fixture labels
and string process-exit field. The next reader stopped on a one-ULP CV rounding
difference between Python versions. Both failures are retained. Only derived
CV fields allow a difference of at most two ULPs; raw and aggregate percentiles,
extrema and gate decisions must still match exactly. This does not change a
rate, scheduler or stability threshold. A further retained reader uses the
explicit maximum aggregation for the worst metric.

Original evidence stays in its earlier immutable seals. The new comparison
receipt, reader, input declarations, exact producer/source/manifest references
and failed attempts are retained in
[evidence/cpp-canonical-matched-review-20261014](evidence/cpp-canonical-matched-review-20261014/retention-seal.json). No historical C++ or
NexusFIX campaign cell is promoted by this review.