Problem

The v1.5.6 release builds failed on macOS (and were one step from the same fate on Windows) with:

ninja: error: '.../ITK-build/lib/libITKCommon-5.4.a', needed by 'GenerateCLP/bin/GenerateCLP',
missing and no known rule to make it

Root cause was a poisoned dependency cache, not anything in the release. The macOS and Windows workflows save the DCMTK/ITK/zlib cache with:

if: always() && steps.<restore>.outputs.cache-hit != 'true'

always() also evaluates true when a job is cancelled. The cancel-in-progress concurrency rule cancels superseded master pushes mid-build, and the cancelled job still ran the save step — archiving a half-built tree. A half-compiled ITK has ITKConfig.cmake (written at configure time) but not its libITKCommon-*.a (written later during compilation), so the archive is a valid cache of a broken tree.

Verified from the step record of a cancelled run (28198907054, job 83532943609): Build dcmqi → cancelled at 20:38:15, Test/Package/Upload → skipped, but Save DCMTK, ITK and ZLIB cache → success at 20:38:18 — the exact created_at of the resulting 57.7 MB partial cache (a complete one is ~785 MB).

It was self-perpetuating: once a partial exists, later runs get a cache hit, which (a) restores the broken tree and (b) satisfies cache-hit != 'true' so the save step is skipped — nothing ever overwrites the bad entry.

Fix

Gate the save on the build step having actually succeeded:

if: ${{ !cancelled() && steps.build.outcome == 'success' && steps.<restore>.outputs.cache-hit != 'true' }}

• Preserves the original intent (the old comment): still caches deps when a later step (test, package) fails — !cancelled() keeps the step eligible after such a failure.
• Never snapshots a tree from a cancelled or failed build — steps.build.outcome == 'success' is the real gate.

Also hardens the macOS Validate restored cache step to require a real ITK library (libITKCommon-*.a), not just ITKConfig.cmake — exactly the gap that let the partial pass validation and then fail at link. Any partial that slips through is now rejected and rebuilt cleanly.

Linux uses the combined actions/cache action, whose post-save does not run on cancellation, so it never exhibited this and is left unchanged.

Notes for reviewer / follow-up

• The poisoned macOS caches were already deleted manually and v1.5.6 was re-run successfully — this PR prevents recurrence.
• A lurking 69.7 MB partial cache under the windows-deps-* key should still be deleted manually (its restore-keys prefix fallback means it can be resurrected on a future Windows run); this PR stops new ones being created.

🤖 Generated with Claude Code