# Classic FIX I/O and deadline progress review — 2026-10-12 Seven reproduced correctness and idle-CPU defects are corrected. The scoped C++, Java, .NET and Rust qualifications pass in development and open editions. The session `codex-review-progress-1012` uses isolated worktrees and `LIBHFT_BUILD_ROOT`. The regression baseline starts at `ca11208ad`, with only the new fixtures overlaid for each language. Builds and suites run serially per host, using CPU leases and a check for foreign heavy jobs. Local parallelism is capped at 16, HP at 4, and Scan at 8. ## Corrections - C++: propagate a negative worker `StepOnce` result through the default `AcceptorThreadPool::Run`. A caller-supplied `RunWith` returning false still requests a clean stop. - C++: preserve a failing port's result in `MultiPortAcceptor::StepOnce`, while stepping the other ports. Work reported by another port cannot cancel out the error, and two errors cannot turn the result into an unrecognized sum. - Java: enforce a worker TLS initiator's Logon deadline while its reader is blocked waiting for the peer. Serialize the timeout with Logon admission and reject admission on a retired channel. - Java: enforce the same worker's local Logout confirmation deadline when the peer remains silent. Preserve the existing stop and callback ownership. - .NET: permit an idle poll when buffered bytes form an incomplete FIX frame. Complete or malformed input remains runnable; TLS read completion retains its existing wake. The extra parser call is confined to the idle-wait path. - Rust: make `advance` and `step_read` stop on a closed transport before reporting a ready frame or processing the buffered tail. Preserve an earlier terminal error and allow a fresh reconnect to establish its own receive state. - Rust: keep a peer's queued Logout response draining before retiring the managed connection. Each acceptor turn still services other peers. A positive LogoutTimeout retains its original deadline; zero retains the existing disabled-timeout meaning. Forced close and store faults retire immediately. The Rust change concerns replies to an admitted peer Logout. It does not change the application-requested Logout or failed-start policies being reviewed separately. ## Results The .NET tests-only baseline reproduced all six incomplete-frame cases: 59–60 owner-thread CPU ticks in each 600 ms interval, across TCP/TLS and header/body/checksum splits. The two deliberate spinning controls and the native-required check passed: 6 failures, 3 passes, no skips. The candidate also checks completion of a 71-frame burst, which exceeds one receive-pass budget, without waiting for the configured 1000 ms idle interval. The corrected Java baseline reproduced only the two worker TLS deadlines: 2 failures, 8 passes, no errors or skips. Its original run had a fixture configuration error (the TLS context's acceptor settings omitted the listener port): 4 passes and 6 setup errors. Those reports are retained as setup errors, not evidence of a runtime defect. Adding the explicit ephemeral port fixed the fixture before the corrected baseline ran. The Rust baseline reproduced five closed-receive failures, with the two sticky-error and reconnect controls passing. Its peer-Logout selection reproduced three failures, with the forced-close/store-fault control passing. The failure assertions show the post-close sequence/store frontier advancing, a ready frame being returned on a closed transport, premature acceptor retirement, and the initiator losing its queued FIX bytes. Exact failing identities and panic text are retained in `rust/scan/baseline-verified.json`. The C++ baseline reproduced four failures with seven passing controls. A fatal poll returned success after a second poll; multi-port results `(-1, -1)` became `-2`, and `(-1, 1)` / `(1, -1)` became success. Both editions now pass all 11 focused cases and the full pool and multi-port registered executables. The complete C++ gate contains 13 verified stages, with assertions enabled. Java passes 75 cases in each edition, including the native-required check, deadline, TLS, callback epoch, worker retirement and stop controls. Its fresh JNI libraries use the corrected, C++-qualified headers at the normal build-root locations. Each exact JUnit inventory has zero failures, errors or skips. | Rust edition | Closed receive | Peer Logout | Session | Initiator | Acceptor | | --- | ---: | ---: | ---: | ---: | ---: | | Development | 7 | 4 | 442 | 37 | 47 | | Open | 7 | 4 | 428 | 31 | 38 | All Rust selections passed with zero failures or ignored cases. The 11 focused cases also appear in the session/owner totals; they are not additional unique coverage. Exact listed and executed inventories match. .NET passes all 114 selected cases in each edition, with exact discovery/TRX identity equality and zero failures or skips. The selection covers idle polling, TLS read/write completion, stop behavior, Logon SendingTime and latency checks, Logout admission, QuickFIX writer ownership and the required native library. All six partial-frame cases record zero owner-thread CPU ticks in 600 ms in each edition; the same-host TCP/TLS spinning controls record 59 and 60 ticks. Every partial-frame case also delivers its completed 71-frame burst within the fixture's 300 ms bound under a 1000 ms idle poll. This is a measured park and wake result, not a claim of zero CPU use at finer clock resolution. This is a scoped correctness and idle-CPU qualification, not a throughput benchmark or a full repository test run. ## Evidence The completed evidence is retained under `~/archive/libhft/reviews/2026-10-12-codex-review-progress-1012/`, including raw TRX/JUnit output, exact test inventories, compiler commands, baseline failures, candidate passes, source/artifact hashes and the upstream integration audit. The local .NET baseline uses SDK 8.0.423/runtime 8.0.29 and GCC 15/C++17 for the native-required library at its normal build-root path. No native library is relocated with an ad hoc build flag. - C++: `cpp/evidence-progress-hp` and `cpp/final-progress-evidence-audit.json`; HP GCC 14.2.1/C++11/O1/LTO1, `-UNDEBUG`, j4. All 1,274 candidate C++ inputs match the integrated local tree. - Java: `evidence/hp-java-deadline-baseline-port`, candidate raw reports and `java/final-deadline-evidence-audit.json`. The audit verifies 4,914 copied evidence files, both exact 75-case inventories, and actual required JNI paths/hashes. Of 2,170 conservative input pins, 2,169 match the integrated tree. The sole difference, `cross_binding_api_inventory.py`, changes a static .NET suppression-pattern check and is used by neither Maven nor the native makefile; the Java results do not qualify that tool change. - Rust: `rust/scan`, `rust/QUALIFICATION.md` and `rust/local-scan-source-equivalence.json`; Scan Rust/Cargo 1.92.0, default features, test debug information disabled, j8, one test thread. All 848 Rust/shared-fixture/dictionary/tool input hashes match locally. This default Rust build has no C++/JNI/.NET native engine-library dependency. - .NET: `dotnet/hp-candidate-evidence-26c841a0`, including both raw TRX reports, discovery inventories, CPU observations in `tests.log` and six successful build/discovery/test stages. HP uses explicit SDK 8.0.131, runtime 8.0.31, GCC 14.2.1/C++17 and j4. Both editions force a fresh native build at the normal build-root path and use fresh managed outputs; the test copy must match the native artifact hash. All 1,823 C++/.NET source pins match the local candidate. Local integration at `26c841a00` imports the upstream .NET SendingTime/Logout correction (`9fd67df1d`) and managed QuickFIX writer ownership (`10c31016b`). Their regression classes are included in the passing .NET candidate. These imports change none of the qualified C++/Java/Rust runtime or test inputs. The final upstream import, `d640e3c55`, adds a separate benchmark driver and its evidence; it changes none of these qualified inputs. The archived `integrated-inputs-after-rebase.json` records the final source equivalence. Local C++ and .NET candidate waits were canceled before any compiler or test launched; the preserved cancellation receipts are not test results.