[BryanFRD]

A release that failed at push left the release commit, tags and a checkpoint local only. The checkpoint was pinned to the pre-commit HEAD, so later runs never matched it and reported nothing to release.

• Checkpoint now stores the planned tags and the floating tags it created.
• A checkpoint is accepted when HEAD is either the pre-commit HEAD or the recorded release commit.
• When planning finds nothing to release and a checkpoint awaits its push at HEAD, the run resumes the push and publish phases with the stored tags (manifest is not rewritten).
• Resumed runs force-push the recorded floating tags.

Closes #1341

[ferrfleet-interne]

One blocking finding: a resumed release skips registry publishers and post-publish hooks, because hook_contexts is never rebuilt, and then reports success. One nit on which plan wins when bumps are detected again. Persisting the checkpoint and floating tags looks correct otherwise.

[ferrfleet-interne]

Both earlier findings are fixed in 7536b33, and their threads are resolved.

Nit: resumed_hook_contexts sets channel: None, but package.rs:253 fills it from the prerelease context. So a resumed prerelease passes hooks and publishers no channel. The checkpoint doesn't record the channel, so fixing this would mean persisting it or re-deriving it from the branch config. That's optional.