Ship parallel engines on every platform at a fixed -j 8 (macOS static libomp, Windows vcomp140 beside the engine) (#1923)

[swapnilpaliwal-sd]

Fixes #1923 — every platform ships PARALLEL (OpenMP) engines, solving at a fixed -j 8, with no switch.

Before this, only Linux shipped parallel: macOS engines were built without OpenMP, and Windows was made serial in #1914 because /openmp needs vcomp140.dll, absent without the Visual C++ Redistributable.

What changes (.github/workflows/build-engines.yml, graph/pipeline/run-souffle.sh, packaging/engine-package.json)

• macOS (arm64, x64): OpenMP with libomp.a (built for macOS 12) linked statically; otool -L shows only /usr/lib/libc++ and libSystem. A guard fails the build if an engine links anything outside /usr/lib and /System.
• Windows: /openmp again, with vcomp140.dll (and whatever VC runtime DLL it imports; none for MSVC 14.44) copied beside each engine, so a machine with no VC++ Redistributable runs it. The guard now requires every DLL an engine (or a shipped DLL) imports to be part of Windows or shipped beside it; the engine package carries the DLLs.
• Linux: unchanged (static libgomp, build-engines: engines start on a clean machine — static libgomp on Linux, no OpenMP on Windows (#1913) #1914).
• Run time: a packaged parallel engine always solves at -j 8; no environment variable changes it. A local source compile keeps its probe and AXIOM_SOLVE_PARALLEL / AXIOMCODE_SOLVE_THREADS.

Validation (packaged engines + packed CLI, npm install -g, clean environment)

platform	full suite	parallel = serial (direct re-solve, 5 runs × java/py/ts/js/c#)	crashes
darwin-arm64 (this Mac, Homebrew/souffle off PATH)	e2e 222/227 0 fail · MCP 40/40 · extras 80/81 0 fail	0 differing relations, every run	0
win32-x64 (stock Windows Server 2022, no VC++ redist, 8 vCPU)	e2e 222/227 0 fail · MCP all pass · extras 80/81 0 fail	0 differing relations, every run	0

Solve time, serial → -j 8 (avg of 5):

project (facts)	macOS arm64	Windows
Python, larger (200 MB)	—	31.4 s → 15.3 s
Java, larger (113 MB)	—	13.8 s → 9.9 s
JavaScript (33 MB)	4.73 → 3.48 s	7.25 → 4.70 s
C# (11 MB)	0.81 → 0.61 s	2.75 → 1.49 s
Python (16 MB)	1.45 → 1.15 s	4.69 → 2.8 s
Java (10 MB)	0.65 → 0.54 s	2.74 → 2.60 s
TypeScript (6 MB)	0.50 → 2.13 s	1.87 → 1.29 s

Known trade-off: on a very small input (6 MB of facts) macOS TypeScript is slower at -j 8 than serial; accepted — the gain is on large repositories.

darwin-x64 (x64 Node under Rosetta on Apple Silicon, clean environment): e2e 222/227, 0 failed · MCP 40/40 · extras 80/81, 0 failed · crashscan 0 crash markers. Parallel = serial: 0 differing relations at index level (java 39, python 95, ts 58, js 51, c# 44 relations) and in all 5 direct parallel re-solves per language; 0 crashes. Timing under Rosetta, with the case suites running on the same machine, so it is noisy: serial build → -j 8 median: java 2.1 → 2.5 s, python 2.7 → 4.2 s, ts 1.8 → 9.3 s, js 7.8 → 8.7 s, c# 2.6 → 1.1 s. On these small inputs -j 8 under translation does not pay; native Intel Macs were not available to measure.

Case suites on this branch (macOS arm64, local compile): java (--oracle --no-torture), typescript (--oracle), python, javascript (--oracle), csharp: all pass. Query cases (tests/run.py): all five languages pass. The first local run had 2 java checks fail in library-auto-discovers-imported-dependencies because no Vineflower decompiler was installed locally (CI installs it). The csharp suite needs the Roslyn oracle built. With both in place everything passes, and engine-package-test.sh passes too. A wrapper around the installed engine confirmed the packaged path passes -j 8 whatever AXIOM_SOLVE_PARALLEL (unset/0/1) or AXIOMCODE_SOLVE_THREADS says.