# HP complete canonical physical matrices — 2026-10-14

One-time named installer registration is now complete on HP and SCAN. Both
hosts verify all fifteen frozen installer entries and all six pending runtime
bundles. Routine installation uses scoped passwordless dispatch. The fixed
bootstrap and its saved terminal-output diagnostics completed successfully.

## Actual HP results

Five separate canonical generations now complete their physical tiny and full
matrices on `hp/sfc-DAC/same-host-netns`. Each holds the host's shared and
isolated exclusive leases, requested with high priority. These are two
processes on one machine connected through the physical Solarflare DAC,
accepted as Stage 5 by this campaign. They are not cross-machine runs.

| Generation | NOS/ER | Market data | Mixed | Tick-to-trade |
|---|---|---|---|---|
| .NET Native canonical | S5 pass | S5 pass | S5 pass | S5 pass |
| C++ CORE canonical | S5 pass | S5 pass | S5 pass | S5 pass |
| Rust open canonical | S5 pass | S5 pass | S5 pass | S5 pass |
| TrueFix current-thread canonical | S5 pass | S5 pass | S5 diag: achieved rate below 98% | S5 pass |
| NanoFIX public-transport canonical | S5 pass | S5 pass | S5 pass | S5 pass |

Each generation retains four tiny pairs and twelve full pairs, with 120,080
primary samples, 120,080 scheduler-delay samples and 330,250 original
application frames per peer. Together: **80 actual pairs, 600,400 primary
samples, 600,400 delay samples and 1,651,250 original frames per peer**.
All production source byte/mode inventories remain unchanged. The installed
runtime and launcher hashes match the previously admitted immutable payloads.
No new SDK build or SDK patch was needed for these runs.

The earlier waiting processes exited without retained results after their
long-lived command transports changed. A resume wrapper now writes each
foreground queue's output to a persistent private log. It skips completed
capsules and refuses to overwrite partial outputs. Every actual run remains
wrapped by the frozen CPU-slot queue. Other sessions' work was preserved.

## Independent replay and portability correction

All ten tiny/full matrices have been copied with original bytes and modes
to BEELINK2 and independently replayed without sockets or SDK compilation.
The reader reproduces every full original body, typed owner/count, raw
primary/delay series, quantile, true extremum, physical counter identity and
polling sequence, and report decision exactly. It also reproduces TrueFix's
mixed-flow failure; that flow receives no report promotion.

The first relocated Native replay passed raw metrics and fixture checks but
failed its final row comparison. The grader still followed the producer's
absolute `artifact_dir` paths. The correction maps only a deep copy of retained
run metadata to the relocated case directory after checking its original
flow/run identity. Original records and frozen producer sources are unchanged.
The original failure is documented separately and eighteen counter/relocation
controls pass. The maintained reader correction is published as `2d13adcd0`.

## Immutable retention

`evidence/hp-complete-canonical-physical-20261014` retains **6,204 files,
962,231,789 bytes**, seal
`05fed686e9b5ceccfc00bea866c7c2844479bb88c761c81c0e15fd6483a248f8`.
It includes every original pair, original grading source, complete frozen
producer source inventory/archive and queue helper, persistent queue output,
all independent receipts and replay commands, and the consumer correction.
6,205 original imported files include the seal; the import receipt is separate.
Earlier seals and their historical grades are unchanged.

## Remaining gates

SCAN's separately queued matrices continue under normal-priority exclusive
CPU leases. Its other sessions' occupied slots are respected. Both hosts'
separate legacy three-engine Rust diagnostics and same-host cold queues
remain queued where occupied; their completion is not claimed here.

The new matched Rust comparison gate (`fba4b0dc6`) passes 24 structural/source
controls and an actual HP source/profile/row preflight. It replays both full
arms before admitting a comparison, binds the exported source inventories,
checks the complete fixture and report contract, and withholds ratios for
blocked flows. The two-rig campaign comparison remains pending SCAN results.
The current matched C++/NexusFIX comparison, Java anchor rate gates, remaining
comparators and whole-harness release admission remain open. These individual
physical passes do not by themselves complete the full plan.