# Current Java callback reply physical outcome — 2026-10-17

SCAN completes the explicit
`java-pure+callback+epsilon-gc+nio-rw+split+reply-cb` configuration from the
unchanged `2d94a107cfaa403a5255254831045bdc3cdb43c5` public SDK/harness.
The actual installed generation runs unprivileged through the registered
DAC entry, with `mlxA/mlxB` physical link counters and actual host facts.
An outer `isol-no-shared` lease excludes shared jobs during timing.

All four tiny pairs and twelve report pairs complete. Both exact cold replays
pass, including original preflight frames, full owner folds, measured admin,
120,000 report primary and 120,000 delay samples, physical counters and
unchanged original inputs. The tiny profile adds 80 observations of each
kind. Report traffic accounts for 330,300 application frames per peer.
Application body files close after the first ten pre-warmup cycles.

| Flow | Achieved cycles/s over three report runs | Retained report blockers |
|---|---:|---|
| NOS/ER | 9,554.970–9,693.368 | Rate |
| MD | 9,213.085–9,390.642 | Rate |
| Mixed | 6,010.900–6,151.918 | Rate and scheduler delay |
| Tick to trade | 9,703.765–9,802.067 | Rate |

Every flow remains **S5 diag**. Completing original samples and functional
proof does not erase these failures. No compact result cell is created or
replaces an earlier passing cell. HP's first physical run of this exact
configuration remains queued at high priority.

## Independent decisions and products

The separately frozen `1b03dca7a` reader passes all 111 reader/archive/merge
controls on SCAN, replays this original report and the original Philadelphia
report, and matches their source, host collector, actual JVM/runtime/CPU,
clock and immutable product proofs. The actual original installed archives
pass their member/container binding. The one-rig merger reproduces a blocked
decision for all four flows, with every comparison metric/ratio withheld.

The exact compact decision is
`private/docs/internal/bench/comparisons/java-callback-reply-scan-20261017.json`,
9,212 bytes, SHA-256
`fdc404b7ec2ee1d638ce86c87dae0dfc02ced23fec497c9704139331fb517505`.
An independent local merge from the exact original report and archive receipts
reproduces those bytes. It declares **one rig**, so it cannot admit the
required two-rig current Java comparison.

Original report SHA-256:
`f96d257565bd9bdb5b462dd9417f4ce1f583cc03662db52e603e180931ef8a99`;
original installed-archive receipt SHA-256:
`c274034d695cf7d79c7479edb64df33cbce0fe0f195328f614768925f82f0a6c`.
Actual cold decisions remain under
`~/build/codex-fullbench/libhft/java-reply-owner-comparison-callback-20261017-v1`
on SCAN. Original physical output remains under
`java-current-sdk-callback-physical-20261017-v1`. Raw series and application
captures stay on the benchmark host.

The same explicit reader independently replays the original poll owner and
again retains all four blocked decisions. Its new one-rig merge is retained
separately; the previously published poll decision is not replaced.

## Complete standalone harness control gate

The actual 933-file standalone benchmark export at `1b03dca7a` passes
**2,250 controls plus 58 subtests**, zero skips/failures, with the two
SDK-source-only guards explicitly deselected. Its inventory SHA-256 is
`e5a36f33460c6ce9cd07c7fc5c3146f04023e3e4827558d116c79fec1b8a77e0`.
Source bytes/modes remain unchanged; no `.git`, SDK engine tree or internal
directory is present. The original completion receipt on SCAN at
`java-reply-owner-portable-20261017-v1/completed.json` has SHA-256
`6b3f3133095474044e123e881d9ad66ed1ef7f13ef7cf3cab9627d329bac007d`.
This gate does not qualify an SDK build, Stage 5 or a public release.

## HP first physical collection completed

The separately qualified original callback product
`6471ba9f2d4ac5698131244829068276e80d819e1df51f756af3ce5a081d07aa`
now completes its first quiet high-priority HP Solarflare DAC collection.
Four tiny and twelve report pairs, both exact independent cold replays,
full owner/admin checks and physical counter proof pass. The report retains
120,000 primary and 120,000 scheduler-delay samples; pre-warmup captures
remain limited to ten cycles. Source and original result bytes are unchanged.

| Flow | Actual cycles/s, min–max | Stage 5 outcome |
|---|---:|---|
| NOS/ER | 9,962.574–9,974.052 | pass |
| MD | 9,849.868–9,957.468 | pass |
| MD + NOS/ER | 9,597.312–9,731.508 | diagnostic: rate |
| Tick to trade | 9,823.502–9,972.835 | pass |

Only the three passing new explicit `+reply-cb` compact keys are added.
Every existing result byte remains intact at this import. Together with
the separate peer replacement import, the catalogue has 658 cells. Mixed
and all original SCAN callback failures stay failed and supply no metric.
The independently replayed two-rig callback decision admits the three HP
matches only; the whole current Java comparison stays blocked. See the
[complete matched decision](RECORD_2026-10-17_java_reply_owner_comparison.md).