feat: Implement Neural Link Transaction/Undo Stack for Agent-Driven UI Mutations

Summary

Implement a transaction stack within the Neural Link server that records all agent-driven UI mutations as reversible operations, enabling undo/redo, atomic batching, and session-persistent audit trails.

A2A Context (Fat Ticket Protocol)

Agent: Claude Opus 4.6 (Antigravity) Session Origin: Multi-Window Agent Shell architecture session

Problem

When agents modify a live UI (creating/removing components, setting properties), there is no mechanism to:

• Undo a mutation ("undo that", "go back")
• Batch related mutations into a single atomic transaction
• Persist the mutation history for session replay
• Audit what an agent changed and when

This is a hard requirement for conversational UIs where non-technical users interact with agents that modify their application in real-time. Users must be able to say "undo that" and have confidence the system will revert cleanly.

Proposed Solution

Implement a transaction stack that:

1. Records every mutation (create_component, remove_component, set_instance_properties) as a reversible operation
2. Supports undo(n) and redo(n) operations via dedicated NL tools
3. Groups related mutations into named transactions (e.g., "add-summary-grid")
4. Optionally persists the stack to the Memory Core for cross-session continuity

Design

// Transaction Model
{
  id:         'tx-001',
  name:       'add-summary-grid',
  timestamp:  '2026-04-10T10:30:00Z',
  operations: [
    {
      type:    'create_component',
      target:  'neo-container-42',
      forward: { ntype: 'grid', columns: [...] },
      reverse: { action: 'remove_component', componentId: 'neo-grid-99' }
    },
    {
      type:    'set_instance_properties',
      target:  'neo-container-42',
      forward: { height: 400 },
      reverse: { height: 300 }
    }
  ],
  status: 'committed' // | 'rolled-back'
}

New Tools Required

Architectural Context

• Foundational for the Multi-Window Agent Shell concept
• Must work across SharedWorker boundaries (all windows share the transaction stack)
• The Memory Core integration means an agent can resume modifications across sessions
• The stack lives in the NL server process, not in the App Worker

Pitfalls Identified

• Property mutations via set_instance_properties need to capture the previous value for the reverse operation — this requires a pre-read before every set
• Component creation reverse operations need the component ID (only known after creation), so forward ops must be awaited before the reverse is finalized
• In SharedWorker mode, the transaction stack must be window-agnostic (a single stack across all windows)

Acceptance Criteria

• Transaction recording for create_component, remove_component, set_instance_properties
• undo and redo tools functional
• Named transaction grouping via begin_transaction/commit_transaction
• E2E test: create component → undo → verify component removed → redo → verify component restored