[reynaldichernando]

closes #123

• Suggestions: a small model now writes the chips after the turn, so the build model no longer holds the composer while it writes them.
• End-of-turn save: the composer no longer waits for the save, which finishes in the background, cutting that gap from 2.2s to under 10ms.
• Manifest check: its reads now run in parallel and skip .puter/, taking it from about 1s to 0.2s.
• Cache clear: it now starts right after the AI writes files instead of when update_preview is called, saving about 1.5 to 2s per edit.