# Native .NET first-CPU physical outcomes

The [separately qualified first-CPU profile](RECORD_2026-10-17_dotnet_native_cpu_placement.md)
completes all 32 scheduled tiny/report pairs on HP and SCAN. HP passes all
four full flows. SCAN passes NOS, MD and tick; mixed fails the unchanged
9,800/s rate floor in one of its three complete runs. All endpoint exits,
fixture/owner/admin/clock/CPU/physical and sample checks pass. No unchanged
timing retry is run.

| Rig | Flow | Full grade | Achieved /s, min–max | Blockers |
|---|---|---|---:|---|
| hp | nos_er | S5 pass | 9893.397–9894.184 | none |
| hp | md | S5 pass | 9887.118–9888.395 | none |
| hp | md_nos_er | S5 pass | 9879.866–9880.395 | none |
| hp | tick_to_trade | S5 pass | 9891.354–9892.722 | none |
| scan | nos_er | S5 pass | 9981.487–9982.108 | none |
| scan | md | S5 pass | 9890.133–9891.067 | none |
| scan | md_nos_er | S5 diag | 9777.110–9812.828 | achieved_rate_below_98pct |
| scan | tick_to_trade | S5 pass | 9978.160–9979.961 | none |

SCAN mixed retains original rates 9,777.110, 9,812.828 and 9,802.215/s. Its
first run fails the floor; the complete row is diagnostic and writes no
compact result or ratio. The three SCAN tick p50s are 25.585, 26.102 and
25.390 microseconds and pass the unchanged full stability gates. This is a
different declared configuration from the inherited-profile run and supplies
no causal claim about the earlier failure.

**The two configurations are not merged into an all-four-flow result.**
Inherited SCAN tick remains failed. First-CPU SCAN mixed remains failed.
Their individual passing cells do not admit a complete SCAN configuration
or a current .NET peer comparison. HP's passes are independent evidence on
that rig only.

## Immutable source and products

- Physical source pin: `c5c1e3577617a9ad4882e59c119f6b41b2128ec9`.
- Complete stripped source: 904 files, inventory SHA-256
  `5cc9844fdb96ab1a1f994c4cd7cc8474dd083b56f29feb54ceb60c7688ad304b`.
- Qualified driver DLL, identical on both hosts:
  `f497d9b795a2e0580ef147d0799c019730727943a74a3f4b70e6ba838299a4d4`.
- Existing native library:
  `91bdc88f8cd2d375262daafdc233ce2b16bc49e8fcae0789da12af999c737dc9`.
- HP installed generation: `761f7c8cd2f29f0702ce08c9830a91518670463cd9096df70b2c222d8c7ef3b0`.
- SCAN installed generation: `ac861dc9a34c44c498cb53cc4175e15161aa8e0dec39d49ca10a51088d362169`.

Seven actual compilation inputs match the qualified source. The immutable
bundle pins the existing managed/native assets, matching SDK source identity,
actual runtime information and original successful qualification receipts.
No SDK/native/root-helper change or new password step is needed. The fixed
launcher supplies the complete budgets; actual foreground startup masks
verify HP server 8 from 8–12/client 13 from 13–17, SCAN server 10 from
10–13/client 14 from 14–17. Every endpoint/run reports the declared profile
and both observed masks. Host probes bind both physical NIC ports, kernel,
firmware, namespaces and budgets. Preparation uses `isol`; physical collection
uses `isol-no-shared`, high priority on HP and normal on SCAN. Slot/process
state and disk are checked before work.

## Independent proof and compact publication

All four tiny and twelve full pairs per host retain exact canonical
preflight/owner/admin observations and complete primary/delay samples. Full
profile is 1,000 warmups, 10,000 measured cycles, three runs per flow, target
10,000 cycles/s. Monotonic RTT ends after final parsing/extraction/checks;
tick ends at public polling handler delivery before the benchmark view.
Probe is integer zero. Capture closes after ten pre-warmup cycles and writes
no measured application frames.

Both full and tiny cold replays pass. Full proof reproduces all 120,000
primary/120,000 delay samples and 330,300 application frames per peer per
host, every original quantile/true extremum and every promotion decision.
Tiny proof retains 80 primary/80 delay samples and 350 frames per peer per
host. Source/report bytes and modes remain unchanged. No reader compiles,
opens a socket or reruns timing.

Seven producer-written compact cells are copied byte-exactly, at most 2,000
CDF points per latency kind. Distinct `+first-cpu` keys preserve every prior
inherited result and failed gate. Original working roots on both hosts:

- `~/.cache/libhft/fullbench-dotnet-native-first-cpu-physical-20261017-v1`
- `~/build/codex-fullbench/libhft/dotnet-native-first-cpu-physical-20261017-v1`

Only admitted compact cells and this record enter Git/website data. No raw
capture, sample stream, log or source tarball enters a new archive. Managed
.NET public-session qualification, peer rate/fixture gates, a complete current
SCAN Native configuration and whole-harness release gates remain open.