# TrueFix borrowed-book callback diagnostic

The Stage 4 builder accepts the optional build-time setting
`TRUEFIX_BENCH_BOOK_DELIVERY=borrowed_callback`. The default remains
`message_clone`. Neither the admitted Rust physical payload nor its hashes change.

This experiment decodes and folds the complete production `35=W` inside
TrueFix's borrowed `Application.from_app` callback. It checks every group entry,
both side counts, size and numeric price, then carries only the correlated
request index, symbol and best ask to the driver. The original full `Message`
clone/channel delivery remains for requests, orders and execution reports.
The production TCP transport, source pin, dictionary overlay, prepared tick
clock, pacing and full canonical exchanges stay the same. An exact-source
feature and its compiled-main hash are recorded with the build inputs.

The HP localhost run uses the current-thread runtime, disjoint CPU masks under
an `isol` lease, and allows concurrent shared conformance work. It is diagnostic
evidence. Five excluded warmups followed by 20 measured cycles pass on all four
flows. The medium profile retains 10,000 primary and scheduler-delay samples
in each of three fresh process pairs per flow; all canonical audits and
functional checks pass.

| Flow | Achieved /s, three runs | Mean p99 µs | Stability | Local rate gate |
|---|---|---:|---|---|
| `nos_er` | 9,990.632; 9,991.627; 9,990.572 | 66.446 | PASS | pass |
| `md` | 9,990.292; 9,990.663; 9,990.925 | 88.041 | PASS | pass |
| `md_nos_er` | 6,995.639; 6,969.568; 7,002.856 | 155.042 | PASS | fail |
| `tick_to_trade` | 9,989.730; 9,990.685; 9,990.804 | 62.691 | PASS | pass |

Mixed flow still fails the offered-rate requirement. The earlier current-thread
diagnostic reached 6,660–6,729/s, but these were different run windows with
shared work present. This is not a controlled speedup claim or a solution to
the full-plan rate gate. The default message-clone mode also builds and passes
four fresh tiny functional/canonical checks with the new source.
Its market-data client extraction sink exactly matches the borrowed callback
mode for the same tiny payload sequence, confirming the complete book fold.

All retained manifest file hashes are verified. The complete inputs, compile
logs, raw series, wire dumps, role observations, process/slot snapshots and
invocations are in
[truefix-borrowed-book-20261002](evidence/truefix-borrowed-book-20261002/).
The experimental binary, dictionary and provenance are kept separately in HP's
`~/.cache/libhft/fullbench-truefix-borrowed-book-20261002`.

## Host queue compatibility correction

The new `host-cpu-slot status` exits 1 when a pool is busy, including a lease
held by the caller. An initial diagnostic wrapper therefore stopped before
compilation. Its failed-attempt log and the later active-slot snapshot are
retained. Informational
status calls now use `|| true`; the actual outermost build/benchmark still
requires `host-cpu-slot wait`. The affected session-owned HP and SCAN physical
watchers were corrected and restarted only while waiting on installer markers,
with sleeping children. Active jobs and other sessions were left running.