[cuihtlauac]

Summary

Lowers the supported OCaml floor from 5.1 to 4.11 for both packages.

The library and its FFI stack were already 4.x-compatible. The only OCaml 5-only use was Domain.recommended_domain_count in the build configurator (config/discover.ml), used to pick the -j count for the vendored CMake fallback. It now shells out to getconf _NPROCESSORS_ONLN (falling back to nproc --all) — both report online CPUs, ignore OMP_NUM_THREADS (pinned to 1 in CI), and are compiler-independent.

4.11 is the oldest compiler still carried by the default opam-repository (everything below it has been moved to the archive repo), so it is the lowest floor that is practically installable and CI-testable today.

Changes

• config/discover.ml — replace Domain.recommended_domain_count with a getconf/nproc shell-out via C.Process.run
• dune-project, xgboost.opam, xgboost-bench.opam — (ocaml (>= 4.11)) (opam files are hand-maintained; generate_opam_files false)
• .github/workflows/ci.yml — matrix oldest compiler 5.1 → 4.11; the oldest (4.11) and newest (5.5) exercise both the system fast path and the vendored CMake fallback, middle versions run the fast path only
• CHANGES.md — changelog entry

Verification

• dune build -p xgboost and opam lint pass locally (OCaml 5.5.1, system libxgboost 3.0.4)
• Scanned src/, bin/, config/: no post-4.08 stdlib or syntax features remain
• The actual 4.11 build (opam solve for a 4.11-compatible ctypes/dune-configurator set, plus the vendored path under 4.11) is validated by CI

🤖 Generated with Claude Code

[cuihtlauac]

Note: this must not be merged-and-released as-is. The opam-repository CI for the previous release (xgboost.0.2.2) is failing, and the root cause is a latent bug in code this PR ships unchanged.

On workers without a system libxgboost >= 3.0, the build falls back to compiling vendored XGBoost from source. config/discover.ml (build_jobs) sizes cmake --build -j N from the online CPU count. On opam-repo-ci's 80-core ocluster workers that means -j 80, which spawns enough parallel g++ to exhaust the job's memory and the kernel OOM-kills cc1plus (surfaces as "cmake build failed (exit 2)").

PR20's own CI is green only because GitHub-hosted runners are ~4 vCPU, so the fallback builds at -j 4 and fits in RAM. It hides the bug rather than disproving it: the next release cut from main will reproduce the same OOM on opam-repo-ci.

Fix needed before merge/release: cap the vendored build parallelism. nproc//proc/meminfo report host figures, not the job budget, so the cap should come from the cgroup limit with a conservative constant fallback (+ an env override).

Full diagnosis and ongoing discussion with @mseri: ocaml/opam-repository#30964 (comment)

[cuihtlauac]

Addressed by 0.2.3 (min 6 ncpus, per @mseri on ocaml/opam-repository#30964); this PR is rebased on it and keeps the cap.