# Explicit Epsilon and synchronous-compilation experiment

The [bounded-capture two-rig matrix](RECORD_2026-10-16_java_preflight_physical.md)
retains rate failures for every LibHFT Java flow and tail-stability failures
for Philadelphia-fast market data. A separate explicit runtime configuration
now tests whether changing compilation scheduling alongside Epsilon GC can
improve those outcomes. This is a hypothesis; no improvement is claimed.

Both canonical JVM builders accept `--jvm-profile epsilon-eager`. It combines
the existing 512 MiB prefaulted Epsilon profile with the existing comparator's
compilation options: `-Xbatch`, `-XX:-TieredCompilation`,
`-XX:CompileThreshold=100` and `-XX:CICompilerCount=2`. Requested warmup, measured
counts, pacing, extraction work, validation and clock spans are unchanged.
The public SDK binaries and existing frozen runtimes are unchanged.

The compact writer matches the longest complete declared argument prefix,
requires the actual Epsilon collector/placement facts, and gives this
configuration a distinct `+epsilon-gc+eager-jit` variant. Missing or altered
compilation options cannot borrow its key. Existing default, Epsilon and
comparator eager launch options and result keys retain their exact meaning.

All 392 focused runtime-profile, preflight and generation controls pass with
zero skips, including negative controls for partial compilation options,
changed threshold, collector and heap. These are SDK-free controller checks.
Actual JVM build, functional/cold qualification and newly pinned quiet
physical reports remain required before this experiment can write results.
No failed original timing is retried under its old variant, no admission
threshold is changed, and no new Stage 5 or comparison grade is claimed here.

## HP build and functional qualification

The complete stripped source at `4b83299c26c27572e7ebd64af51bbce3b89aa51f`
has 886 files and inventory SHA-256
`fbd4335c5b782f6b7c133fe29daad4f526b028ff91f4d8fd95b6eb034f17540d`.
Both drivers build against their original unpatched SDK JARs under HP's
high-priority `isol` lease, with shared conformance work concurrent. Both
pass twelve actual functional pairs and complete independent cold replay,
with zero skips. LibHFT refuses 107 corrupted inputs and Philadelphia 89.
Actual endpoint arguments and collector observations bind the new keys.

| Driver | Adapter JAR SHA-256 | Original SDK JAR SHA-256 |
|---|---|---|
| LibHFT callback, NIO read/write, split | `4c92180329b88cd6edc790cdfade55336b272b0590136691a7c15545457602e2` | `d459e3f66ae7f098b4715d7044636dc22c670c18a3da0b088a2a12affab6d330` |
| Philadelphia-fast callback, split | `8b85f82732db22c33d8e19f576bfe7b036576bad99e28d9cb53ab363a4ad7120` | `396136037b0f41e635e3aea233c3f31f4a388318b445f4afe1609b9a6d28a7b7` |

The frozen harness inventory stays unchanged. Qualified products are staged
for both physical rigs; preparation and report execution use separate leases.
The working qualification root is
`~/build/codex-fullbench/libhft/java-epsilon-eager-qualification-20261017-v1`
on HP. Quiet physical execution and any timing/matched admission remain
pending; the original failed configurations are not replaced.

## SCAN physical outcome

The quiet physical matrix completes both drivers' four tiny and twelve
report pairs on SCAN's Mellanox DAC. All endpoint exits, canonical body and
owner checks, namespace/CPU admission and physical wire-counter gates pass.
Each report retains 120,000 primary and 120,000 scheduler-delay samples,
with 330,300 application frames per peer. Independent cold replay reproduces
all eight diagnostic decisions exactly and leaves originals unchanged.
The frozen source inventory remains unchanged.

Every report flow misses the required 9,800/s achieved rate. The observed
minimum–maximum rates across three runs are:

| Flow | LibHFT Java /s | Philadelphia-fast /s | Additional LibHFT blockers |
|---|---:|---:|---|
| NOS/ER | 7,464–7,511 | 8,706–8,734 | percentile stability |
| MD | 5,279–7,141 | 8,932–8,946 | percentile stability, scheduler delay |
| MD + NOS/ER | 1,546–4,798 | 8,513–8,515 | percentile stability, scheduler delay |
| Tick to trade | 5,180–5,845 | 7,635–7,652 | scheduler delay |

Philadelphia has no additional blocker. These failed cells create no compact
results, replace no admitted result and establish no matched comparison.
The experiment is not retried unchanged. HP's distinct physical collection
remains queued at high priority behind shared conformance work.

The immutable generations are
`28493b630b74d6e309589a7b34d7cbc1de2d39b63504ebbbe2c778852a821c11`
(LibHFT) and
`56bb8edc9beb7874de1066676c6da6de131a7d26ed773989dd903b823535d206`
(Philadelphia). Original working output is under
`~/build/codex-fullbench/libhft/java-epsilon-eager-physical-20261017-v1`
on SCAN. No raw output is copied into Git or the website.

## HP follow-up completed

The queued high-priority HP matrix now completes under a quiet lease. Its
original source/runtime remains unchanged; actual outcomes, admitted compact
cells and retained failures are in the [HP follow-up](RECORD_2026-10-17_hp_canonical_followup.md).
This appendix supersedes the earlier pending-HP scheduling status, while
preserving the original SCAN outcomes and campaign-specific comparison scope.