# Canonical .NET direct byte formatter physical outcomes

The [qualified formatter](RECORD_2026-10-17_dotnet_byte_bodies.md) completes
96 actual physical tiny/report pairs: three separately declared configurations
on each accepted DAC rig. Native first-CPU passes all four flows on both rigs.
Native inherited and actual ManagedSession inherited pass all four on HP.
SCAN Native inherited NOS and Managed inherited tick retain p50 stability
failures. Every row passes all other functional, physical and timing gates.

All 22 successful producer-written compact cells are imported byte-exactly.
Four keys are newly admitted (Native inherited SCAN tick, Native first-CPU
SCAN mixed, Managed mixed on each rig); eighteen successful existing keys
receive their latest collection. The two failed cells do not replace their
earlier good results. A set of retained cells from different generations does
not make this collection pass all four flows. No configuration union or .NET
peer comparison is admitted.

| Rig | Configuration | Flow | Full grade | Achieved cycles/s, min–max | Blockers |
|---|---|---|---|---:|---|
| hp | native-inherit | nos_er | S5 pass | 9987.687–9987.812 | none |
| hp | native-inherit | md | S5 pass | 9987.617–9987.785 | none |
| hp | native-inherit | md_nos_er | S5 pass | 9987.161–9987.379 | none |
| hp | native-inherit | tick_to_trade | S5 pass | 9987.150–9987.420 | none |
| hp | native-first | nos_er | S5 pass | 9987.257–9987.927 | none |
| hp | native-first | md | S5 pass | 9987.360–9987.737 | none |
| hp | native-first | md_nos_er | S5 pass | 9986.961–9987.061 | none |
| hp | native-first | tick_to_trade | S5 pass | 9987.363–9987.910 | none |
| hp | managed-inherit | nos_er | S5 pass | 9983.501–9984.052 | none |
| hp | managed-inherit | md | S5 pass | 9916.948–9917.319 | none |
| hp | managed-inherit | md_nos_er | S5 pass | 9915.618–9916.179 | none |
| hp | managed-inherit | tick_to_trade | S5 pass | 9981.033–9981.353 | none |
| scan | native-inherit | nos_er | S5 diag | 9981.799–9982.409 | full_percentile_stability |
| scan | native-inherit | md | S5 pass | 9981.321–9981.903 | none |
| scan | native-inherit | md_nos_er | S5 pass | 9974.387–9979.936 | none |
| scan | native-inherit | tick_to_trade | S5 pass | 9978.283–9980.684 | none |
| scan | native-first | nos_er | S5 pass | 9981.564–9982.332 | none |
| scan | native-first | md | S5 pass | 9980.567–9981.486 | none |
| scan | native-first | md_nos_er | S5 pass | 9977.370–9978.134 | none |
| scan | native-first | tick_to_trade | S5 pass | 9980.441–9981.886 | none |
| scan | managed-inherit | nos_er | S5 pass | 9976.467–9976.755 | none |
| scan | managed-inherit | md | S5 pass | 9975.602–9976.005 | none |
| scan | managed-inherit | md_nos_er | S5 pass | 9805.861–9823.247 | none |
| scan | managed-inherit | tick_to_trade | S5 diag | 9975.358–9975.528 | full_percentile_stability |

SCAN Native inherited NOS p50s are 28.661, 23.862 and 25.572 µs: sample CV
9.343610% exceeds the unchanged 5% bound. Managed inherited tick p50s are
34.610, 31.183 and 33.273 µs: sample CV 5.230552% also exceeds 5%. Both
p99 CVs remain below 10% and p99.9 CVs below 25%. Rate passes for every new
row, including Managed SCAN mixed at 9,823.247, 9,820.578 and 9,805.861/s.
Older rate/stability failures remain in their original records. No unchanged
failed timing retry is made and no causal allocation/CPU claim is inferred.

## Original production sessions and immutable source

- Physical source: `38932a0e9b59f2bf09c8d9d3221b422e914c60c7`.
- Complete stripped source: 909 files, inventory SHA-256
  `9e97fe17d0f4c49b8277e20976af8194453a12678f6631a10d8e71a151a2967c`.
- Native DLL: `b730a18e484b833c664a587f2e88b919ea444cffe9e3d72c5398f7b38155050a`.
- Managed DLL: `47133bbce59459b5769ae46e6fc6a59d0546f24999adef176e972be06daa3770`.
- Unchanged native library:
  `91bdc88f8cd2d375262daafdc233ce2b16bc49e8fcae0789da12af999c737dc9`.
- Unpatched SDK:
  `uncommitted-functional-source-26310f12584fd0f392af231da8c2ea10067f50bb337db118cc79a2895da54386`.

| Rig | Configuration | Immutable installed generation |
|---|---|---|
| hp | native-inherit | `1272bec6a754a217e3f71f525ec673484e545c8791586c01fa2e99572742a02c` |
| hp | native-first | `19cdd0bce904653787f52e1875c0c0a811318c25ccdbc5e33f9e8355a232c57f` |
| hp | managed-inherit | `92931f39aa7300c570126774e26f973ba4f557d6eff4180ee48dc16783615599` |
| scan | native-inherit | `bd9c38ceaf855459ab46c71c949a671444c3c376ef1b93f35ba6ab9586f6f0d4` |
| scan | native-first | `09f01c272e3b1fa55eba5bf764800a87fe1f445b680eeb7a16dd1393665d9a0a` |
| scan | managed-inherit | `fb04d421fc24135023fcaee10930ada24d115523f13691ddce5ca7dbd0f149c6` |

The seven actual compilation inputs of each consumer match its qualified
source. Runtime files match the original functional manifests and the same
DLLs on both hosts. Native uses public NativeEngine/NativeSession batches and
PollMessage; managed uses public ManagedSession/IFastHandler and the pre-send
gate, with no native engine loaded. The SDKs own session state, parsing,
serialization, actual writes, Logon and Logout. The corrected declared managed
listener bind remains explicit. No root helper, sudo policy, SDK/native source,
clock, required field/extraction or grading threshold changes.

HP uses Solarflare server CPUs 8–12/client 13–17; SCAN Mellanox server
10–13/client 14–17. First-CPU foreground observations are 8/13 and 10/14,
within the unchanged full budgets. Other configurations keep their actual
inherited masks. Host probes bind actual kernel/NIC/firmware/namespace/CPU
facts. Preparation holds `isol`, timing holds `isol-no-shared`, with high
priority on HP and normal priority on SCAN; both slot/process and free-disk
states are checked before work. Distinct configured ports cover 63200–63215,
63300–63315 and 63400–63415 respectively. Applications run as the benchmark
user; only the existing fixed helpers use root.

## Complete clocks, capture and independent cold replay

Each report retains 1,000 warmups, 10,000 measured cycles, three runs per flow
and target 10,000 cycles/s. Complete RTT ends after final required extraction,
folding and checks; tick starts after SDK serialization before actual write
and ends at public polling-handler return or managed callback entry. All
process-local clocks remain monotonic. Boundary probe is integer zero.

Only ten pre-warmup cycles are captured, then streams close before warmup.
Measured capture/trace is zero. All actual bilateral frame, callback, prepared
and completed write, owner sink/fold, admin, raw count and physical counter
checks pass, including both diagnostics. Each host/configuration's tiny and
full readers independently reproduce original quantiles, true extrema and
every promotion decision. Full replay has 120,000 primary/120,000 delay
samples and 330,300 application frames per peer per configuration/host; tiny
has 80/80 and 350. Across all six collections, full proof covers 720,000
primary/720,000 delay samples and 1,981,800 application frames per peer.
Original bytes/modes remain unchanged; no replay compiles, opens sockets or
repeats physical timing. All 1,018 focused checks additionally pass from this
complete stripped archive, without skips. Export closure/scrub findings are
zero; whole-harness release remains separate.

Both hosts retain the original source and working roots:

- `~/.cache/libhft/fullbench-dotnet-byte-bodies-physical-20261017-v1`
- `~/build/codex-fullbench/libhft/dotnet-byte-bodies-{native-inherit,native-first,managed-inherit}-physical-20261017-v1`

Only compact results with at most 2,000 CDF points per latency kind and this
summary enter Git/website data. Raw captures/samples, logs and source archives
stay on the hosts. Managed first-CPU is the next separately declared profile:
it must pass both actual ordinary/probe qualifications before new complete
physical matrices. The inherited tick failure is retained. Current peer and
whole-harness release gates remain open.