feat: run dependency graphs

[privatenumber]

Problem

Publishing a dependency closure requires running per-package preparation in dependency-first order: leaves pack first, each package receives its direct dependencies' results, and independent packages overlap. Without a shared runner, every orchestration layer would reimplement readiness tracking, result relay, and failure gating — and re-derive the same concurrency and error semantics slightly differently each time. For example, a naive sequential loop would turn the diamond case (core, two adapters, app) into four serialized steps, while an ad-hoc concurrent version risks running a dependent before its dependencies resolve or losing track of which result belongs to which package.

Changes

runDependencyGraph() takes generic nodes and a value-producing callback, validates duplicates, unknown keys, and cycles before invoking anything, then schedules leaves-first with independent nodes overlapping. Each callback receives only its direct dependencies' results, every node runs exactly once, and results return keyed by node. A callback failure rejects without invoking blocked dependents, while already-running independent work settles cleanly. The runner knows nothing about packages or Git — adapting PublishGraphNode targets into it is explicitly left for the manifest-rewriting stage.

[github-actions]

📊 Package size report   No changes

File	Before	After
Total (Includes all files)	79.4 kB	79.4 kB
Tarball size	29.0 kB	29.0 kB

Unchanged files

File	Size
dist/index.js	71.1 kB
LICENSE	1.1 kB
package.json	1.5 kB
README.md	5.7 kB

_{🤖 This report was automatically generated by pkg-size-action}