Refactor: emit skills.toml / skills.lock via serde instead of string templates

[jon-devlapaz]

Summary

manifest::format_manifest_entry / format_lock_entry / quote hand-build TOML. Prefer serde (de)serialization with the same types used for load, so schema and emit cannot drift.

Why

• Unknown-fields deny on read but write is ad hoc
• Easy to forget a field when extending lock format
• Testing round-trips becomes natural (lock → bytes → load_lock)

Acceptance

• Written files remain version=1 and human-readable enough for git diffs (pretty/stable key order)
• Round-trip test: generate → parse → equal semantic content
• No required CLI flag changes

Priority

Low — quality of life when next extending the manifest schema.