fix: Ensure mock adapter returns packages with package_id in all code paths

[bokelley]

Summary

Fixed bug where mock adapter and create_media_buy tool were improperly handling pending/async responses, causing confusion about media buy state and returning fake IDs to clients.

Problem

The mock adapter was returning fake media_buy_id values like "pending_question_..." or "pending_step_..." for async/pending operations. This is wrong because:

1. The media buy doesn't exist yet - it's being processed
2. Fake IDs confuse clients - they look real but aren't queryable
3. Violates A2A protocol design - pending operations should return minimal domain data with protocol envelope indicating status

Additionally, media_buy_create.py was trying to save non-existent media buys to the database and process packages even when the operation wasn't complete.

Root Cause Analysis

Architecture: AdCP protocol uses two layers:

1. Domain layer (CreateMediaBuyResponse) - Business data
2. Protocol layer (ProtocolEnvelope) - Wraps with status field

For pending/async operations:

• Domain response should have media_buy_id=None (doesn't exist yet)
• Domain response should have packages=[] (not created yet)
• Protocol envelope provides status="working" or "input-required"
• Protocol envelope provides task_id (workflow_step_id) for tracking

The client uses task_id to poll or wait for webhook, NOT a fake media_buy_id.

Solution (3 commits)

Commit 1 (1d42301): Initial defensive fix - ensure mock adapter returns packages with package_id

• Added package population in async/question paths
• Prevents immediate error but doesn't fix architecture issue

Commit 2 (924c408): Skip package processing for pending responses

• Detects pending by checking media_buy_id.startswith("pending_") or None
• Skips package merge/validation for pending
• Proper architectural separation

Commit 3 (aaaea9c): Remove fake media_buy_id from pending responses

• Mock adapter returns media_buy_id=None for async/question scenarios
• Mock adapter returns packages=[] for pending
• Tool logic handles media_buy_id=None properly
• Database persistence skipped for pending (nothing to save yet)

How It Works Now

For completed operations:

{
  "status": "completed",
  "payload": {
    "media_buy_id": "mb_abc123",
    "packages": [{...}]
  }
}

For pending/async operations:

{
  "status": "working",
  "task_id": "workflow_step_xyz",
  "message": "Processing your media buy...",
  "payload": {
    "buyer_ref": "buyer_123",
    "media_buy_id": null,
    "packages": []
  }
}

Client uses task_id to track, not a fake media_buy_id.

Testing

• Unit tests pass
• Syntax validated
• Code inspection confirms proper handling
• Pre-commit hooks pass

Impact

• Fixes production error on wonderstruck.sales-agent.scope3.com
• Removes confusing fake IDs from client responses
• Properly aligns with AdCP protocol envelope architecture
• No breaking changes for clients (they should check status field anyway)
• Backwards compatible with existing code

🤖 Generated with Claude Code

[bokelley]

Closing in favor of clean PR without schema files. See new PR.