# Fixer-rs: both physical DAC campaigns and compact result admission

## Outcome

Both new configuration campaigns are complete. Each host passes sixteen
original SDK pairs: four small functional runs and three 10,000-sample report
runs per scenario, with 1,000 warmup cycles and ten captured preflight cycles.
Both fresh cold readers replay all 120,080 primary and 120,080 delay values
per host, including functional runs. All original body, owner, phase, clock,
namespace, effective CPU, normal-user, process-exit and physical wire gates
pass. No SDK is rebuilt for physical collection.

HP admits two scenario cells. The remaining six report cells stay diagnostic
under the unchanged 10,000/s contract. The table gives median run percentiles
and achieved aggregate rates; diagnostic rows are explicitly unadmitted.

| Host | Scenario | p50 µs | p99 µs | Achieved cycles/s | Admission |
| --- | --- | ---: | ---: | ---: | --- |
| HP | NOS/ER | 58.547 | 62.495 | 9,982.89 | S5 pass |
| HP | MD | 176.555 | 183.225 | 5,598.46 | rate failure |
| HP | mixed | 248.623 | 261.084 | 3,980.85 | rate and scheduler-delay failures |
| HP | tick-to-trade | 48.983 | 52.348 | 9,982.19 | S5 pass |
| SCAN | NOS/ER | 96.697 | 119.356 | 9,769.18 | rate failure |
| SCAN | MD | 281.202 | 294.359 | 3,511.77 | rate and scheduler-delay failures |
| SCAN | mixed | 399.325 | 418.276 | 2,467.36 | rate and scheduler-delay failures |
| SCAN | tick-to-trade | 87.566 | 99.120 | 9,109.08 | rate failure |

The rate gate applies to every repeat, with a 98% minimum; it does not round
the SCAN NOS/ER result into a pass. No unchanged timing rerun is pending.
The minimum existing Rust comparison and the legacy fixer-rs adapter keep
their separate identities and grades. This is no four-flow matched admission
for the new fixer-rs configuration.

## Configuration and physical evidence

The [preceding timed qualification](RECORD_2026-10-23_fixer_rs_timed_session.md)
pins the original SDK, serializer, complete RTT and callback tick clocks,
first-CPU entry and correct public Session Logout procedure. The same
qualified executable is installed through the previously registered immutable
generation helper. No root helper, sudo policy, registration input or SDK
file changes.

- HP uses Solarflare `sfcB`/server and `sfcA`/client, effective CPUs 8 and 13.
- SCAN uses mlx5 `mlxB`/server and `mlxA`/client, effective CPUs 10 and 14.
- Each report sitting holds both CPU leases through `isol-no-shared`. HP
  requests high priority. Both recorded status and actual inherited lease
  descriptors are checked; running processes and memory are captured first.
- The fixed root entry chooses the namespace, inherited CPU budget and
  normal user. The measured executable records its actual namespace inode,
  inherited/effective CPUs and executable path. The cold reader checks them
  against the fixed entry and observed namespace records.
- Original physical counter snapshots advance on both ports. All asynchronous
  post-run polling attempts remain retained; no probe traffic is fabricated.
- Original host/NIC facts are observed before and after each campaign under
  the benchmark identity. Kernel, physical PCI NIC and firmware agree across
  endpoints and phases. The counter source hash is unchanged.

## Independent result admission

The separate admission process recomputes every original phase/clock/raw
difference, all canonical preflight bodies and measured owner folds, and all
per-run percentiles and report blockers. HP refuses ten actual metadata
mutations plus two well-formed but false empirical CDF points. SCAN refuses
the same ten metadata mutations and admits no cells. Source identity,
quiet-lease, counter, borrowed-engine, namespace/CPU and promoted-report
mutations are refused semantically without altering original files.

The two committed HP cells use flavour
`fixer-rs+custom-order+memory-store+callback-tick+complete-rtt+first-cpu`, method
`dac-canonical-preflight-v1`, kernel TCP, paced 10,000/s. Each retains all
30,000 samples per latency kind through its summaries and a 2,000-point
empirical CDF, plus bounded actual preflight text and source/rig evidence
hashes. Their total is 242,580 bytes. Full raw and clock series remain on the
producing hosts. Website generation publishes summaries and excludes these
per-test CDF arrays from hosting.

The focused source suite passes 217 tests and 24 refusal subtests. The result
format checker passes all 1,228 current source cells. Existing result keys and
the benchmark archive are preserved.

## Exact pins

| Item | Pin or SHA-256 |
| --- | --- |
| Qualified producer | `ed54fbb0df48726cc7405a606f27a1e6e6ff488c` |
| Physical controller/stager | `95ccfdaa98` |
| Cold reader/writer/admission worker | `3f479760d28e736b29b5b64de710b641c7c20192` |
| Shared executable | `fbe59557a908817f85504ad33a8c9ce2b6f6d4b25eb5a146b00520c5cd4401db` |
| HP physical input manifest | `97d15ce3a1ef6466acb85c2de6068dfce72ce8d3f376fb44edf06b96cd9aa216` |
| SCAN physical input manifest | `c97c2dce20a990b5f11aa64562899ec0bcaa0b7b3c20924accb52bcb2d8cdfbb` |
| HP admission input manifest | `8f3b814fd2b32ed342f4d8c48a0c0390697c9ba162a1e02ae1c1516e5232db3f` |
| SCAN admission input manifest | `7117e450b2ebe47adb9296b94aaff0bce0cc20eec6d48824389a2c908ce80556` |
| HP completed campaign | `ef09981c7d33278f7b96f453ab2f14dc87df27f7135e0edcf99cdb207949d7c5` |
| SCAN completed campaign | `8696e126070ba5ec3033e244023de0fff3548857cc578748fff3a8d3db1be300` |
| HP independent replay | `b8b693ad05e2c79b494d787f4616233518e07043d4d6c0bb592c60a4ac894ed9` |
| SCAN independent replay | `5d29dfbed30db4c28c92a199bc776ff177c7b8b2cc5dea7e3469bb7cec078391` |
| HP result admission | `6aea0b9b96751b8bf328f21b1cf3f771eb6a3c9a0001403cec9b5d6720b0ee1b` |
| SCAN result admission | `7ae1143e291dc3237ff172c276fdc34e5a1ac5e1b9ae71280c3912d21ffe8ad9` |

Host working roots are
`~/build/codex-fullbench/libhft/fixer-rs-physical-qualification-20261023-v1`;
bounded local reviews use `fixer-rs-physical-review-20261023-v1/{hp,scan}`.
Physical services are `codex-fullbench-fixer-rs-physical-20261023-v1`.
HP's admission uses `codex-fullbench-fixer-rs-admission-20261023-v1`;
SCAN's queued shared reader was replaced by
`codex-fullbench-fixer-rs-admission-isol-20261023-v1`, a normal-priority pure
replay on free isolated CPUs alongside shared work. It repeats no benchmark.
Both final admission services finish successfully with no live own process.
No output is copied into `~/archive`, and no run folder is manually deleted.