AI Preview tf.ai Refactor

Goal

Replace the low-level AI preview snapshot helper surface with tf.ai.* lifecycle transforms, keep the full-document snapshot strategy private, migrate insert-mode preview callers, and preserve the current no-history preview behavior.

Checklist

• [completed] Gather current AI preview/history patterns and relevant learnings
• [completed] Add or update tests for the new preview lifecycle contract
• [completed] Implement tf.ai.* preview transforms in BaseAIPlugin
• [completed] Migrate AI chat preview callsites off direct snapshot helpers
• [completed] Run required verification for packages/ai and apps/www

Findings

• The current preview storage lives in packages/ai/src/lib/transforms/aiStreamSnapshot.ts as a WeakMap keyed by editor.
• The current low-level helper surface is used directly by acceptAIChat, undoAI, resetAIChat, the editor AI kit, and the streaming integration test.
• The existing solution doc already established the correct behavior boundary: preview chunks stay out of history; accept commits one fresh batch; AI undo restores the pre-stream value while preview is active.
• AIPlugin already owns editor mutation semantics like insertNodes, removeMarks, removeNodes, and undo, so preview lifecycle belongs there too.
• The cleanest public surface here is tf.ai.*, while library internals and stricter callsites can still reach the same transforms through editor.getTransforms(BaseAIPlugin).ai where generic editor typing is narrower.
• Package-scoped verification passed for @platejs/ai; filtered apps/www typecheck still has unrelated workspace export/type failures after the required root pnpm build, so that debt remains outside this refactor.