# CORE bounded-capture physical collection

## Inputs and placement

This sitting uses the complete 885-file stripped harness at
`5c2671618f051e0263ac661e048cfbe36b1ed946`, inventory SHA-256
`d478134c164ae41eddff040f362554008227a55a74e994dc737c66a7bfa02857`.
The [qualified callback and receive binaries](RECORD_2026-10-16_cpp_preflight.md)
and their unchanged public SDK are reused at their recorded hashes. Compile
input inventories bind those same native source/header/build bytes in this
controller snapshot; no SDK patch or fresh native compilation is hidden.

Both hosts prepare immutable generations through the existing registered
installer and fixed normal-user namespace entry. Neither root helper changes.
HP prepares under a high-priority `isol` lease; its timing request remains
queued at high priority until the shared conformance suite releases the host.
Only this session's waiting timing requests are reordered to let functional
checks and preparation use isolated CPUs alongside shared work. No running
benchmark or another session's queue/process is interrupted.

SCAN executes under `isol-no-shared`, with Mellanox physical DAC namespaces
`mlxA/mlxB`, server CPUs `10-13` and client CPUs `14-17`. The unprivileged
probe binds actual UID, namespace inode and CPU placement, kernel
`5.14.0-687.44.1.el9_8.x86_64`, PCI model `0x15b3:0x1015`, driver
`mlx5_core` and firmware `14.32.1912 (MT_2420110034)`. Physical packet-counter
updates, root-entry admission, original preflight bodies, complete owner
folds/counts, administration and both endpoint exits pass on every pair.

| Variant | Immutable generation |
|---|---|
| Callback CORE | `2f7cd1faa6ea2db2a08b72803cdd030e2bfba80dc70f13c1aa975001a5a6682b` |
| Receive-clock CORE | `e8923a6264c1703a18ab8190a4ac34e9317fb17ce9c0cf45f02faa471664f0bd` |

## SCAN decisions

Each variant runs four tiny fixture pairs, followed by twelve report pairs:
1,000 warmup cycles, 10,000 measured samples per run, three runs per flow and
10,000 cycles/s target. Ten functional cycles precede warmup. Only their
application bytes enter the closed directional capture streams; all live
counts and measured owner folds remain complete.

| Variant | Flow | Achieved rates /s | Report decision |
|---|---|---|---|
| Callback | NOS/ER | 9991.45, 9990.98, 9990.90 | Pass |
| Callback | MD | 9991.46, 9991.41, 9991.58 | Pass |
| Callback | Mixed | 9991.05, 9989.69, 9989.09 | Pass |
| Callback | Tick-to-trade | 9987.69, 9991.24, 9991.55 | Pass |
| Receive | NOS/ER | 9992.08, 9991.64, 9990.09 | Pass |
| Receive | MD | 9991.12, 9991.11, 9991.40 | Pass |
| Receive | Mixed | 9989.21, 9989.08, 9989.87 | Pass |
| Receive | Tick-to-trade | 9991.07, 9990.45, 9991.40 | **Blocked: percentile stability** |

The receive tick's per-run p99 values are 34.144, 28.946 and 28.690 microseconds;
their sample CV exceeds the 10% gate. Rate success does not override that
failure. This variant gets three compact cells and no complete four-flow
admission; the failed cell does not replace any stored result or acquire a
ratio. Callback CORE passes all four individual flows on this rig.

Both independent frozen-policy cold readers reproduce every original
quantile, true maximum and decision exactly. Each report has 120,000 primary
samples, 120,000 delay samples and 330,300 application frames per peer.
Including both tiny matrices, SCAN completes 32 actual pairs, 240,160 primary
and delay samples, and 660,950 application frames per peer. Capture contains
only 400 preflight frames per peer. Source bytes/modes remain unchanged.

The seven admitted compact JSON files are imported byte-exactly under
`private/benchmarks/retained/results/scan/dac-canonical-preflight-v1/`, with separate
`cpp+core+callback` and `cpp+core+receive-clock` keys. Their distributions pool
all measured samples, percentiles retain run medians and maximum is the
largest actual run extremum. Every stored maximum equals its CDF endpoint.
The repository-wide result/key check passes. It first caught an import
root that duplicated the result path; only those seven new temporary copies
were removed and the original bytes were placed at the proper keys before
commit. No raw capture, samples, log, executable or source archive is committed.

Working output is
`~/build/codex-fullbench/libhft/cpp-preflight-physical-20261016-v1` on SCAN.
HP's same-method collection, the peer's bounded-capture migration and any new
two-rig matched comparison remain pending. The older admitted C++ comparison
keeps its original method and evidence.

## HP follow-up completed

The queued high-priority HP matrix now completes under a quiet lease. Its
original source/runtime remains unchanged; actual outcomes, admitted compact
cells and retained failures are in the [HP follow-up](RECORD_2026-10-17_hp_canonical_followup.md).
This appendix supersedes the earlier pending-HP scheduling status, while
preserving the original SCAN outcomes and campaign-specific comparison scope.