# Java bounded-capture physical collection

## Frozen inputs and admission

The complete stripped harness is frozen at
`208e972a2fd460f300882d4af8977afdd852b16c`, with source inventory SHA-256
`321aebffa246bc411e34182afdc3d179c325ead86c6330bbf6a4eb0a539a3190`.
Its 882 source/template files remain byte- and mode-identical after HP's
collection. Staging verifies every Java compile-input digest against this
source and preserves the unpatched SDK hashes from the
[preflight qualification](RECORD_2026-10-16_java_preflight_compact_results.md).

Both immutable runtimes use Epsilon GC and split JVM thread placement. Java
also explicitly selects the NIO read/write socket profile. Their different
variants remain visible in result identities. The existing generic generation
installer and fixed namespace entry suffice; no further sudo registration,
new privileged executable or manual installation is needed.

The unprivileged provenance app runs through that same fixed entry on both
DAC endpoints, during preparation and again immediately before traffic. Its
observed UID/GID is 1000. CPU masks and namespace inodes bind to the actual
entry namespaces. The two HP ports report `sfc`, PCI `0x1924:0x0a03`, firmware
`6.4.2.1020 rx1 tx1`, and kernel `5.14.0-687.44.1.el9_8.x86_64`. These facts
are retained in compact results; no hardware or firmware value is guessed
or borrowed from another host.

## HP physical result

Preparation holds HP's high-priority `isol` lease. The full foreground
campaign holds high-priority `isol-no-shared`; other sessions queue normally.
Each engine completes four tiny physical pairs and twelve full physical pairs.
Every pair passes fixture, original capture, fixed entry and wire-counter
admission with successful endpoint exits. Each full flow retains three runs
of 10,000 primary samples and 10,000 scheduler-delay samples, after 1,000
warmup and ten pre-warmup functional cycles per run.

The frozen independent cold reader reproduces all original grades, quantiles
and true maxima without changing source/output bytes or opening sockets.
For each engine it verifies 120,000 primary samples, 120,000 delay samples and
330,300 application frames per peer. Captures contain only the preflight
cycles; measured accounting and independent owner sinks continue in process.

| Engine / flow | Actual rates across the three runs (/s) | Timing decision |
|---|---|---|
| Java NOS/ER | 9819.41, 8236.71, 9913.66 | rate, stability and delay fail |
| Java MD | 9914.55, 9898.54, 9632.93 | rate and stability fail |
| Java mixed | 7170.77, 7682.94, 7966.82 | rate, stability and delay fail |
| Java tick-to-trade | 9772.75, 9873.12, 9810.87 | rate fail |
| Philadelphia-fast NOS/ER | 9990.35, 9987.31, 9991.17 | S5 pass |
| Philadelphia-fast MD | 9983.81, 9810.25, 9990.72 | stability fail |
| Philadelphia-fast mixed | 9983.67, 9981.09, 9981.40 | S5 pass |
| Philadelphia-fast tick-to-trade | 9988.11, 9988.90, 9990.58 | S5 pass |

Only the three passing Philadelphia cells are imported under
`private/benchmarks/retained/results/hp/dac-canonical-preflight-v1/philadelphia+fast+callback+epsilon-gc+split/kernel/`.
Each retains the run-median percentiles, largest maximum, pooled CDF capped at
2,000 points per kind and the graded ten-cycle pre-warmup exchange. Failed
cells write no compact file and replace no prior result. These cells do not
establish a four-flow or matched Java comparison. No failed timing is retried
without a separate reason or declared source/configuration change.

SCAN has identical source/products; its collection completed subsequently as
recorded below. Host working
output is under `~/build/codex-fullbench/libhft/preflight-physical-20261016-v1`;
raw output is neither committed nor copied into a new archive.

## Earlier local matrices and the aggregate-count correction

HP's separately frozen source at `63cd1c6252f0277f8cad4f4c6e5d9274836e4e07`
completes both Epsilon callback local matrices and original-policy cold
replays. Java misses rate on all four flows and additional delay/stability
gates; Philadelphia-fast passes all four local timing gates. These are
localhost diagnostics and confer no physical admission.

Their pair receipts correctly include the preflight cycles, but both local
manifest producers still hardcode 330,000 application frames per peer and
their original replay omits checking that aggregate. The correct sum is
330,300 per engine. Historical manifests and their original readers are
unchanged; the original cold success does not certify that summary field.
The fresh physical producer/reader already sums actual receipts and is
unaffected.

Both local controllers now derive the three aggregate counts from audited
pair receipts. Cold replay refuses missing, altered, Boolean or legacy
hardcoded counts. The regression covers the actual ten-cycle preflight delta
and exercises both replay entrypoints after pair audit. This changes metadata
validation without rerunning timing or rewriting immutable outputs.

All 85 focused aggregate, compact-result and host-provenance checks pass with
zero skips. Every one of the 559 compact results passes the format checker.
Export readiness selects 881 files with zero private/unresolved references,
dead documentation links or scrub findings; 66 documented source-root escapes
remain a separate standalone qualification gate.

## Remaining work

Complete current-method Java comparison, the remaining language collectors
and whole-harness build/licence/release gates remain open. The website source
now has a complete compact-result catalogue; its new sitting-record projection
is in progress. The historical website archive is preserved; no live deployment
is attempted.

## SCAN collection and two-rig outcome

SCAN's normal-priority `isol` preparation succeeds through the existing
registration. Its foreground physical run then holds both CPU pools under
`isol-no-shared`, starting at 22:51 UTC. All sixteen actual pairs per engine
pass fixture, fixed-entry, original capture and physical-counter admission.
The exact frozen independent reader reproduces all full report decisions,
quantiles and extrema without altering the original source or outputs.
Per engine, it verifies twelve full pairs, 120,000 primary samples, 120,000
delay samples and 330,300 application frames per peer.

The observed SCAN NIC is `mlx5_core`, PCI `0x15b3:0x1015`, firmware
`14.32.1912 (MT_2420110034)`, under kernel `5.14.0-687.44.1.el9_8.x86_64`.
The unprivileged probe again binds both namespace inodes and CPU budgets;
the compact cells retain its original observations and digests.

| Engine / flow | Actual rates across the three runs (/s) | Timing decision |
|---|---|---|
| Java NOS/ER | 8917.55, 8987.71, 9395.08 | rate and stability fail |
| Java MD | 8442.24, 7300.37, 8684.87 | rate, stability and delay fail |
| Java mixed | 4803.05, 4911.34, 4900.87 | rate and delay fail |
| Java tick-to-trade | 9414.66, 9153.86, 2897.53 | rate, stability and delay fail |
| Philadelphia-fast NOS/ER | 9981.08, 9980.42, 9982.75 | S5 pass |
| Philadelphia-fast MD | 9978.13, 9979.06, 9981.22 | stability fail |
| Philadelphia-fast mixed | 9690.98, 9699.65, 9645.76 | rate fail |
| Philadelphia-fast tick-to-trade | 9982.10, 9979.86, 9979.89 | S5 pass |

Only the NOS/ER and tick-to-trade SCAN cells enter
`private/benchmarks/retained/results/scan/dac-canonical-preflight-v1/philadelphia+fast+callback+epsilon-gc+split/kernel/`.
Together the two rigs complete 64 actual physical pairs and admit five
individual Philadelphia cells. All LibHFT Java timing cells and all eight
matched flow comparisons remain blocked. No failed cell replaces a stored
result, no unchanged timing is retried, and no ratio or complete comparison
is published from these partial passes. Every compact cell passes the result
format checker. Raw working output stays on its original host.