Low-latency lab: contracts, counterexamples and measurements

C++20 / x86-64. This is a set of research artifacts, not a production trading engine. Four articles under _posts/2026-09-22-* explain the contracts and separate implemented behavior from proposed extensions.

Build

Linux / Windows with CMake and a C++20 toolchain:

cmake -S . -B build-release -DCMAKE_BUILD_TYPE=Release
cmake --build build-release --parallel
ctest --test-dir build-release --output-on-failure

On multi-config generators use --config Release and ctest -C Release. Tests use always-enabled checks; they do not disappear under NDEBUG.

This session used Zig 0.15.2’s Clang 20.1.2 directly on Windows:

zig c++ -std=c++20 -O3 -Wall -Wextra -Wpedantic -pthread -Iinclude tests/research.cpp -o build-zig/research.exe
zig c++ -std=c++20 -O3 -Wall -Wextra -Wpedantic -pthread -Iinclude tests/stress.cpp -o build-zig/stress.exe
zig c++ -std=c++20 -O3 -Wall -Wextra -Wpedantic -pthread -Iinclude src/trading_loop.cpp -o build-zig/trading_loop.exe
zig c++ -std=c++20 -O3 -Wall -Wextra -Wpedantic -pthread -Iinclude src/benchmark.cpp -o build-zig/benchmark.exe
zig c++ -std=c++20 -O1 -g -fsanitize=undefined -fno-sanitize-recover=all -pthread -Iinclude tests/research.cpp -o build-zig/research-ubsan.exe

Create build-zig before these commands. The Linux-only journal was cross-compiled:

zig c++ -target x86_64-linux-gnu -std=c++20 -O3 -Wall -Wextra -Wpedantic -pthread -Iinclude src/journal_demo.cpp -o build-zig/journal_demo

CMake is supplied for reproduction; it was not available on this session’s host. Direct compiler builds and Windows tests were actually run. Cross-compilation is not a Linux runtime test.

Evidence and boundaries

Artifact	Actual claim
`include/ll/spsc_queue.hpp`	SPSC only; nothrow element operations; acquire/release publication and slot recycling; safe destruction requires quiescence
`include/ll/bounded_mpmc_queue.hpp`	Bounded no-mutex ring; false means next slot unavailable; does not implement a strict-empty lock-free FIFO contract
`tests/research.cpp`	Deterministic reservation hole; full/empty and counter-wrap boundaries; book quarantine/snapshot behavior; seeded differential book; byte-level journal fault injection
`tests/stress.cpp`	One million SPSC ordered messages and 4P/4C exact set verification
`src/trading_loop.cpp`	Synthetic single-stream pipeline, not transport recovery/risk/session implementation
`include/ll/journal_format.hpp`	Versioned little-endian bytes, CRC32C, offline valid-prefix recovery
`src/journal_demo.cpp`	Linux single-writer mmap with MS_SYNC worker, separate published/durable watermarks; runtime untested here
`src/benchmark.cpp`	Same-origin serialized-TSC RTT; checked topology/affinity; equal-quota throughput with mutex comparison

Queue-owned storage is allocation-free; arbitrary T may still allocate or block. Finite MPMC tags require bounded observation age. The book never silently drops deep levels: capacity failure latches quarantine.

Benchmark

./build-release/benchmark all
./build-release/benchmark spsc
./build-release/benchmark mpmc
bash scripts/perf.sh

RTT is two SPSC queues with one request in flight. It includes a conservative CPUID/RDTSC/CPUID timestamp framework; the empty framework is separately printed. Do not divide RTT by two or subtract percentiles to claim single-call latency. The implementation is x86-specific; the Windows path supports one processor group. CPU affinity errors abort the experiment.

The throughput workload uses 16-byte id/inverse messages and 1024 usable slots. Each consumer completes a fixed quota; no per-message shared completion counter. Thread-local checksums are verified after joins. They supplement, not replace, the stress test’s exact set check. Trials are fixed-work and short; no claim of open-loop saturation, fairness or stable maximum arrival rate is made.

The final benchmark ran without concurrent builds/tests dispatched by this task. The desktop OS remained active; no core isolation, fixed-frequency guarantee, scheduler traces or PMU evidence is claimed. See results/RESEARCH.md.

Journal (Linux runtime required)

./build-release/journal_demo create /tmp/unique-journal.bin
./build-release/journal_demo recover /tmp/unique-journal.bin

create refuses existing files (O_EXCL). Use a new path for another run. append returns a published sequence, not a durable acknowledgement. Explicit close joins the worker, reports sync errors, and returns the durable watermark. A destructor alone does not report persistence failures.

The scanner reads a quiescent file and stops at the first invalid/truncated record. It does not skip corruption, perform append-reopen, authenticate data, or prove power-loss atomicity. The Windows tests mutate byte streams; they are not filesystem crash tests.

No live exchange connection, order submission, credentials or account data are used. Network feed recovery, per-subscriber broadcast, order admission epochs, risk reservations and venue reconciliation are analyzed in the articles but not implemented in the synthetic pipeline.