# Registered canonical generations on SCAN's physical DAC

The two fixed generation helpers are now registered on HP and SCAN. Both
installed runtime generations pass their hash/identity checks without password
prompts. Their payloads execute as `yann`, in the fixed physical namespaces and
CPU masks. No runtime or root program is reinstalled for controller changes.

## Original startup failures and corrected controller

The first starts on both hosts fail before benchmark sockets because Python's
child process closes the inherited CPU-slot descriptors. A real-child regression
reproduces that failure and verifies explicit descriptor forwarding. The second
starts fail before benchmark sockets because `host-cpu-slot status` returns 1
for a busy slot, including the controller's own lease. Its output and exit code
are now retained; only codes other than 0/1 abort observation collection. Real
status-child tests cover all three cases. Rig-verification probes precede the
second failure and are not counted as measured benchmark pairs.

All failed logs and partial originals remain intact. The corrected complete
855-file source archive is frozen at `376d5ce9b33789f523b77d77cdfb393d9d3a0224`,
with inventory SHA-256
`eecef3385b27f76bae894d06ce6bc722b4e30c8c50299d18c4b31388a2fb292d`.
Both hosts independently pass 23 source/refusal/descriptor/status checks with
zero skips and unchanged source bytes/modes. Existing runtime generations and
frozen producer archives are untouched. Host status, processes and memory are
checked before work; timed collection uses an outer `isol-no-shared` lease.

## Actual SCAN physical matrices

Each engine completes four tiny and twelve full pairs. Together the 32 actual
pairs retain 240,160 primary samples, the same number of scheduler-delay samples
and 660,500 original application frames per peer. Every original body, typed
owner/fold count, administration, complete sample count, fixed namespace/CPU/
normal-user entry and physical counter/polling check passes. There is no new
compile. The declared topology is `scan/mlx5-DAC/same-host-netns`.

| Engine | NOS/ER | MD | Mixed | Tick-to-trade |
|---|---|---|---|---|
| callback-clock CORE | S5 pass | S5 pass | S5 pass | S5 pass |
| Philadelphia-fast canonical | S5 diag | S5 diag | S5 diag | S5 diag |

Philadelphia retains achieved-rate and scheduler-delay failures on every full
flow. NOS, MD and mixed also retain full stability failures. Its tiny rows retain
rate failures. No threshold, warmup profile or original result is changed, and
no favorable repeat is selected.

All four original profile cold receipts pass. Four additional relocated
SDK-free replays agree exactly with the original receipts, including all
quantiles, true maxima, packet-counter polling and each failed or passing
report decision. Original bytes and modes and reader source remain unchanged.

## Independent C++ comparison on this rig

A shared-slot cold reader independently replays the complete new CORE baseline
and the unchanged original NexusFIX physical report. Profiles, fixture bodies,
owner/counts, administration, primary/delay observations, physical counters,
quiet leases, frozen source/completion/configuration bindings and endpoint CPU
masks agree. Both declare the prepared-send/callback-entry tick span. All four
matched flows pass on SCAN. Percentiles are medians of three run percentiles;
worst is the maximum of all run extrema. The report retains every metric,
including flows where NexusFIX is faster.

This is **one rig**, with `required_rigs=1`. It does not admit the two-rig C++
minimum comparison. HP's actual matrix is queued at high priority behind another
session's shared conformance work; its matched reader waits without holding a
CPU lease. The earlier before-parser CORE/NexusFIX tick mismatch remains an
unchanged blocked historical comparison. Java matching and whole-harness
release gates remain open.

The new seal contains 1,840 retained inputs totalling 372,454,865 bytes. Its
SHA-256 is `c04c40cfcf02ca6a03e92497132772f0ff2380d7bd522763b57a1d5eb2b684cd`.

[Full original matrices, same-host and relocated cold receipts, matched review,
frozen producer, original peer bindings and startup failures](evidence/registered-generation-scan-physical-20261014/retention-seal.json)
are retained separately from earlier generations.