# QuickFIX/J public direct-event configuration

2026-10-12. This is a separately qualified configuration experiment for
the [full-bench plan](../../../docs/bench/PLAN_FULL_BENCH.md), preserving
the default canonical generation and every earlier result.

## Public SDK extension

Private nested subclasses of the stock socket acceptor/initiator override
the protected event-handling extension and supply the public
`EventHandlingStrategy` interface. Incoming messages go to the unchanged
production `Session.next(message)` directly on the MINA event thread.
Failures propagate to the driver. The strategy reports an empty queue
because it dispatches inline; stock connector lifecycle and session timers
remain in use. There is no SDK source patch, reflection or hand-written
protocol response. Original capture, exact canonical bodies, public prepared
responder clocks, actual administration and full measured owners remain.
Observations explicitly declare `direct_public_event_handling_strategy`.

The complete 805-file source inventory is
`experiment-quickfixj-canonical-direct-events-e3f6061c400b2b0b6d0a8f485ff2c3cd3bee86de716c499fca568611bbef88b4`.
The six stock JARs and every runtime product are independently SHA-bound.
Both hosts freshly compile the five selected driver sources and pass
twelve actual pairs, including zero/repeated warmup. Each cohort retains
240 primary/delay samples and 700 original application frames per peer.
Source/products/dependencies remain unchanged; these are functional checks.

## HP full local matrix

A high-priority exclusive lease completes twelve full pairs at 1,000
excluded warmups, 10,000 measured cycles and a 10,000/s offer. Every complete
fixture and owner fold passes. The cohort retains **120,000 primary/delay
samples and 330,000 original application frames per peer**.

| Flow | Achieved cycles/s across three runs | Timing result |
|---|---:|---|
| NOS/ER | 3,914–4,031 | rate/delay fail |
| MD | 3,190–3,322 | rate/delay fail |
| Mixed | 2,093–2,121 | rate/delay fail |
| Tick | 3,867–3,967 | rate/delay fail |

There is no full stability failure in this cohort. The independent default
reader reproduces all twelve receipts, every raw quantile/extremum and
every failed gate exactly, with original bytes/modes unchanged and no new
sockets or compile. Direct events improve some medians, but do not resolve
the report-rate/delay gates. No maintained default, physical result,
historical grade or public comparison is promoted.

## Immutable retention

[Import](evidence/quickfixj-canonical-direct-events-20261012/import-receipt.json)
and [seal](evidence/quickfixj-canonical-direct-events-20261012/retention-seal.json)
retain **530 files / 177,910,734 bytes**, seal SHA-256
`1f442b5adc42e717f7f6bf6964783551465e4fdd5a384128d801a17d8386b17d`.
The supplement contains 36 new actual pairs, 120,480 primary/delay samples
and 331,400 original frames per peer, all exact runtime/source inventories,
and the independent full replay. Both host copies verify the originals.

## Next placement diagnostic

HP's recorded kernel command line isolates CPUs 8–17 with `isolcpus=domain`.
[Linux CPU isolation documentation](https://docs.kernel.org/admin-guide/cpu-isolation.html)
states that these CPUs are removed from scheduler load balancing. An
explicit thread-placement experiment is therefore testing the unchanged
default SDK/driver inside the same four CPUs per endpoint. Its observer
targets only its own JVM threads and records every affinity change.
This is a placement hypothesis; the direct-event results alone do not
establish that thread contention caused their rate failures. No original
cohort is overwritten.