StartOS

What's Changed

Fixed

• make no longer fails in a git repository that has no index yet. s9pk.mk listed $(GIT_DIR)/index as an unconditional prerequisite of the s9pk targets, but git init does not create an index until the first git add — so make aborted with No rule to make target '.git/index', needed by '<id>_x86_64.s9pk' before the pack step ever ran. This hit every freshly scaffolded package: s9pk init-package runs git init and stages nothing, and the generated TODO.md sends the packager straight to make as their first build, while the workflow guide tells them to iterate with a dirty tree and commit once at the end. The existing $(if $(GIT_DIR),…) guard asked only whether a repo exists, not whether these files do. GIT_DEPS now filters through $(wildcard …), so a missing HEAD or index drops out of the prerequisite list instead of halting the build. Nothing else needed to change: s9pk pack already handles a commit-less repo, reporting No git commit found in . — building without a commit hash in the manifest and packing with a null gitHash. These entries are rebuild triggers, not build requirements

Fixed

• ExVer range operations no longer ignore the downstream revision. compareVersionRangePoints and adjacentVersionRangePoints compared the upstream version twice instead of comparing the downstream on the second pass, so two points that shared an upstream but differed in downstream (1.0.0:3 vs 1.0.0:15) collapsed into a single point. Everything built on the truth tables inherited the error — normalize() silently dropped the lower of the two, and intersects() / satisfiable() could answer on a merged point. =1.0.0:0 || =1.0.0:1 normalized to =1.0.0:1.

  The visible consequence was in packed manifests: canMigrateFrom / canMigrateTo are derived from the version graph and normalized, so any package declaring an other version sharing current's upstream advertised a range narrower than the truth (mempool at 3.3.1:15 with other: [3.3.1:3] shipped canMigrateFrom: <=3.3.1:3 rather than <=3.3.1:15). No upgrade actually broke — StartOS resolves migrations through the version graph rather than gating on this field, and the registry index does not yet populate sourceVersion from it — but the manifests were wrong and would have become load-bearing the moment either changed. The Rust implementation derives its point ordering and was never affected.