feat(sleep): adopt reviewed skill subsets safely

[Bogdan (Dan) Baciu (bogdanbaciu21)]

Summary

Supersedes #189, which could not be reopened after its branch was rebased.

Adds an explicit, auditable way to adopt a reviewed subset of multi-skill Sleep proposals. Staging remains non-mutating: live skills change only when adopt_skills() is explicitly called.

What changed

• Adds staged_skills() to inspect staged skill proposals.
• Adds adopt_skills(staging_dir, skill_names) for subset or all-skill adoption.
• Validates every selection before writing anything.
• Backs up existing files and replaces them atomically.
• Rolls back the entire selection if any write fails.
• Records SHA-256 before/after hashes and backup paths.
• Preserves existing file permissions.
• Documents the staging layout, adoption flow, receipts, and migration behavior.

Safety contract

• No implicit adoption.
• Unknown, duplicate, malformed, or unsafe selections fail before writes.
• Unselected skills remain untouched.
• Partial adoption cannot survive a failed write.
• Every adopted change produces a verifiable receipt.

Compatibility

The existing single-skill path is unchanged:

• Manifests without skills remain valid.
• adopt() retains its current behavior.
• adopt_skills() is an independent, opt-in API.
• skill_names=None adopts all staged skills; [] adopts none.

Review follow-up

Addresses all feedback from #189:

• Validates the manifest root.
• Rejects unsafe proposal filenames.
• Preserves file modes during atomic replacement.
• Removes the redundant live-file read.

Validation

• 39 focused tests passed.
• 233 Sleep tests passed, 3 skipped.
• 945 full-suite tests passed, 8 skipped.
• Ruff passed.

Non-goals

• Automatic adoption.
• Changes to gate or scoring behavior.
• Changes to legacy single-skill staging semantics.