Skip to content

ci: migrate rebound/munkres/vmprof to needs.setup.outputs.matrix - #4919

Open
luhenry wants to merge 1 commit into
mainfrom
fix/rebound-munkres-vmprof-matrix
Open

luhenry wants to merge 1 commit into
mainfrom
fix/rebound-munkres-vmprof-matrix

Conversation

@luhenry

@luhenry luhenry commented Nov 18, 2026 •

Copy link
Copy Markdown
Member

Supersedes the now-closed PRs #4911/#4912/#4913, which each widened a hard-coded python/libc matrix to add the missing interpreters those packages' docs/packages/<pkg>.yaml already declared. That fixed the symptom but left the structural duplication in place (a second, independently-maintained interpreter list that can drift out of sync with the YAML again).

This instead converts all three build_wheels jobs to include: ${{ fromJSON(needs.setup.outputs.matrix) }}, the pattern most other build-<pkg>.yml files already use, which is generated directly from each package's own docs/packages/<pkg>.yaml files: entries - eliminating the possibility of this class of bug for these three going forward.

rebound and munkres each layer a per-libc include: override on the old hard-coded matrix (test environment/requires/command). Since the generated matrix carries no custom fields, these overrides are preserved as:

  • munkres: a single test_requires string, converted to an inline matrix.platform == 'musllinux' && ... || ... ternary.
  • rebound: three fields including a multi-line test command with a $(python -c ...) substitution - moved into a per-platform GITHUB_ENV heredoc step instead of an inline ternary, to avoid fragile quote-escaping in a GHA expression. Verified byte-for-byte against the original values via a local bash simulation of both branches (including that the musllinux command's $(python -c ...) text survives literally rather than being expanded early).

vmprof had no override to preserve, just a stale python list and a hard-coded manylinux literal.

Verified each via pending_versions.py --python-matrix --python-exclude cp39 <pkg> against its own docs/packages/<pkg>.yaml:

All matrix.libc references renamed to matrix.platform throughout (job names, CIBW_BUILD, artifact names) to match the generated matrix's field name.

All three still hard-coded their own python/libc matrix in build_wheels,
duplicating docs/packages/<pkg>.yaml and drifting out of sync with it -
the exact bug already fixed once for each of these three in now-closed
PRs #4911/#4912/#4913 (stale lists missing cp310/cp311/cp315). Converting
to needs.setup.outputs.matrix (the pattern most other build-<pkg>.yml
files already use) fixes the coverage gap structurally instead of by
widening a second, independently-maintained list that can drift again.

rebound and munkres each layer a per-libc include: override (test
environment/requires/command) on top of the hard-coded list; since the
generated matrix entries carry no custom fields, rebound's three
overrides move into a per-platform GITHUB_ENV heredoc step (verified
byte-for-byte against the original values via a local bash simulation,
including that the musllinux test-command's $(python -c ...) survives
as literal text rather than being expanded early) and munkres's single
test_requires override becomes an inline matrix.platform ternary.
vmprof had no override to preserve, just a stale python list and a
hard-coded 'manylinux' literal.

Verified each via 'pending_versions.py --python-matrix' against its own
docs/packages/<pkg>.yaml before and after: vmprof now covers cp310-cp315
(all manylinux, matching the closed #4913), munkres covers cp310-cp314
across both platforms (matching #4912), rebound covers cp310-cp314t
across both platforms (matching #4911) - all libc/matrix.libc references
renamed to platform/matrix.platform to match the generated matrix's
field name throughout.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant