Ask for a wheel only on the leg that builds one

[wdolinar]

What broke

xmsconan 2.22.0 made build.py exit 1 when --wheel-dir yields no complete
set of wheels (c614bc5, "Stop publishing a wheel chosen by package-id hash").
The GitHub CI template, however, still passes --wheel-dir wheelhouse on
every matrix leg — and [matrix].pybind_build_types defaults to
["Release"], so a Debug leg has no pybind package to extract.

Every GitHub consumer's Debug legs therefore turn red the moment they float up
to 2.22.x:

 >> All configurations built successfully.
 >> No pybind package found to extract.
error: --wheel-dir wheelhouse was given but no complete set of wheels was extracted

Already observed on master in the wild:

Repo	When	Pattern
xmsgrid	2026-08-21	5 Debug fail / 8 Release pass
xmsmesher	2026-08-22	4 Debug fail / 5 Release pass

xmsinterp and xmsextractor float too and last ran 2026-08-18, before 2.22.0 —
they break on their next run. xmscore never passes --wheel-dir to build.py;
xmsvtk is pinned ==2.16.0; xmsconstraint's GitLab job carries no build_type
filter, so its Release pybind configuration is always present. All three are
unaffected.

The fix

Gate the flag on matrix.build_type == 'Release' — how repair, artifact upload
and deploy in the same workflow are already gated, and what generate_ci
already assumes when it rejects a GitHub build.toml whose pybind_build_types
excludes Release (ci_file_generator.py:210-225).

The guard in build.py is deliberately left alone. --wheel-dir meaning "I
require wheels" is the intended contract; the defect was asking for them on a
leg that cannot produce any. A library that names Debug in
pybind_build_types loses nothing — its Debug wheel was built and discarded
before, since every downstream step was Release-gated already.

The Windows --skip-dependency-libs companion moves inside the gate with it:
build.py reads it only within the if args.wheel_dir branch, so leaving it
outside would put a flag on the Debug command line that reads as staging
control and does nothing.

GitLab is unaffected and unchanged.

Verification

• flake8 . → exit 0 (with flake8-docstrings, -bugbear, -import-order,
  pep8-naming, -tidy-imports, matching CI)
• pytest tests/ -q → 1978 passed, 5 skipped
• Two new tests, both mutation-proven: un-gating --wheel-dir fails
  test_github_build_step_requests_a_wheel_only_on_release[mac], and moving
  --skip-dependency-libs outside the gate fails its own test. The first is
  asserted per job so a regression names the platform.
• Regenerated xmsgrid's and xmsmesher's workflows end-to-end; YAML parses and
  all four build steps carry the gate.

Docs

docs/USAGE.md §10 step 3 and a new §10.1 bullet; rationale added to the
template's header comment.

Ordering — please read before releasing

xmsmesher/master/build.toml still carries has_test_files, which the
KNOWN_KEYS whitelist merged in #112 now rejects. Verified against this
branch:

Error: build.toml has unknown top-level key(s) has_test_files.

Aquaveo/xmsmesher#88 must merge before the release that carries this, or
xmsmesher's CI dies one step earlier, at Generate.

Consumers also need xmsconan ci re-run and the result committed — the
workflow is generated, not floated.