resdata: Add version 6.3.3

[luhenry]

• Package: resdata
• Version: 6.3.3
• Source: https://github.com/equinor/resdata
• Docs: https://github.com/equinor/resdata

Compiles the libresdata C/C++ core and its pybind11 extensions, with backward-cpp, fmt and elfutils built from source by Conan. Upstream publishes no riscv64 wheel.

Mirrors upstream's testing.yml.

Differs from upstream

• build-frontend is build - build[uv] needs a host uv the runner lacks.
• before-build also stages the Conan dependencies' licence texts.
• gpl_sources job archives the elfutils tarball and Conan recipe for the release.
• CIBW_ENVIRONMENT pins SETUPTOOLS_SCM_PRETEND_VERSION - the Apply patches step's git apply dirties the checkout, so git describe reports a post-6.3.3 dev version instead of the clean tag (gotcha 31). The bare env var is needed, not the _FOR_RESDATA suffix: setup.py calls setuptools_scm.get_version() directly without a dist_name, so the per-package variant is never checked.

Matrix: cp312/cp313/cp314 - upstream ships no free-threaded wheel.

Testing

• same as upstream, with numpy, scipy, pandas and pydantic-core taken as wheels only, since the registry trails PyPI for them.
• 15-minute per-test timeout (thread method) - a hang in native code otherwise holds the runner for the whole job.
• test_performance_with_many_keys gets 60 s instead of upstream's 10 s (patch 0002).

License: Wheel statically links elfutils (GPL-2.0-or-later or LGPL-3.0-or-later), backward-cpp, fmt, pybind11, zlib, bzip2, xz and zstd; upstream ships no licence text for them, so the build adds it.

Patches

• 0001-Fix-infinite-loops-at-EOF-where-char-is-unsigned.patch - Backport [https://github.com/equinor/resdata/commit/ce341a3666855c57839466d9daa876252b54c8c0]. Without it test_broken_file loops forever, since char is unsigned on riscv64. Reproduces on x86_64 with -funsigned-char.
• 0002-Raise-test_performance_with_many_keys-timeout-for-sl.patch - Inappropriate [the 10s ceiling is sized for upstream's x86 CI]. The benchmark always does 7 summary.x calls over 5000 keys, about 1.4 s each on riscv64 against 0.1 s on x86_64. The 10 s marker ran out on the seventh call, and since its timeout takes precedence over --timeout=900, it ended the whole session. The work is linear in the key count; nothing loops.

All three build legs (cp312/cp313/cp314) pass on real riscv64 hardware.

[luhenry]

Same mechanic as the other recent campaign PRs: check_commit_messages (job 116023092412) failed on check_no_ai_attribution against the pull_request "opened" webhook's frozen body snapshot, captured before an auto-injected attribution footer was stripped from the PR body. The live PR body has been clean since right after creation.

pr-checks.yml already triggers on edited for exactly this case, and that run has already passed. No fix to push.

[luhenry]

Failure: Build resdata 6.3.3 cp314-manylinux_riscv64 (run 38711493152, head 192bfe1) died in CIBW_BEFORE_BUILD, not in resdata's own build: while conan install was building ninja (a build tool it pulls in), cmake's compiler sanity check links with -static-libstdc++ -static-libgcc and failed with /usr/bin/ld: cannot find -lstdc++.

Root cause: this is riscv64-image-specific, not a cp314/pybind11/ABI issue, and not an upstream resdata bug. quay.io/pypa/manylinux_2_39_riscv64 is Rocky 10-based and only ships the shared libstdc++.so, not the static libstdc++.a that -static-libstdc++ needs. It happens before any Python-version-specific code runs (during dependency staging, same for every cpXYZ leg), so cp312/cp313 would hit the identical failure once their legs reached this step — they just hadn't run yet (this runner only had capacity to start cp314 first). This is exactly gotcha 77 (skills/python-project-porting/references/gotchas/native-deps-and-linking.md), already hit and fixed the same way for this same image in build-ddtrace.yml, build-libclang.yml, and build-coacd.yml.

Fix: added CIBW_BEFORE_ALL_LINUX: dnf -y install libstdc++-static to build-resdata.yml (CRB is already enabled on this image, so no --enablerepo flag needed, matching build-coacd.yml). Pushed as e5aecc5.

I couldn't run a full riscv64 rehearsal locally (this sandbox's Docker can't reach quay.io through its network policy), so validation here is the YAML parsing cleanly plus this being the identical, already-proven fix for the identical error/image elsewhere in the repo. Re-running CI on this PR should confirm it for all three Python legs.