fix(prompt_deploy): normalize Foundry definition to dict for SDK 2.x compatibility

[placerda]

Summary

Fixes a hard failure in prompt_deploy stage that fires when the user's prompt differs from the seed (the created action path) and the runner installs azure-ai-projects >= 2.x.

Symptom hit during a live tutorial recording at step 14 of tutorial-prompt-agent-quickstart.md:

azure.core.exceptions.HttpResponseError: (invalid_payload) required:
Required properties ["kind"] are not present

Root cause

Two compounding issues in src/agentops/pipeline/prompt_deploy.py:

1. _copy_definition returned the wrong shape under SDK 2.x. It called .copy() on the typed PromptAgentDefinition returned by get_version. In SDK 1.x that preserved the typed model so the body serialized as a flat {"kind": "prompt", "model": ..., "instructions": ...}. In SDK 2.x the same .copy() returns a stripped base Model whose JSON shape is {"_data": {"kind": "prompt", ...}}.
2. _create_agent_version wrote kind at the body root via definition.get("kind"). With the post-.copy() stripped model that returns None. The new Foundry API also treats kind strictly as the polymorphic discriminator inside definition, so a root-level kind is at best ignored.

Combined effect — the JSON that reached the service looked like:

{
  "kind": null,
  "definition": {"_data": {"kind": "prompt", "model": "...", "instructions": "..."}},
  "metadata": {...},
  "description": "..."
}

…which the service rejected as malformed.

Why this only fired on one path

Only the created action path runs _copy_definition + _create_agent_version with the typed SDK model:

Path	Definition source	Affected?
reused	uses existing current directly, no create_version call	No
bootstrapped	builds a plain dict from prompt_agent_bootstrap	No
created	round-trips a typed PromptAgentDefinition through .copy()	Yes

That matches what the user saw: the first deploy (where the prompt hadn't changed) returned action: reused and succeeded; the second deploy after they edited the prompt blew up.

Fix

In src/agentops/pipeline/prompt_deploy.py:

• New _definition_to_dict helper normalizes any SDK definition object (dict, typed model with _data, mapping with .items(), or anything with as_dict()) to a plain dict.
• _copy_definition now always returns a deep-copied plain dict.
• _create_agent_version drops the root-level kind from the body and sends a clean {definition, metadata, description} shape.

Verified locally with SDK 2.2.0: serialization of a real PromptAgentDefinition through the fixed path produces the expected flat shape with kind inside definition.

Tests

• Added test_copy_definition_returns_plain_dict_from_sdk_typed_model — regression test that simulates the SDK 2.x .copy() semantics and asserts the result is a flat dict without _data.
• Added test_create_agent_version_body_uses_flat_definition_dict — asserts the body sent to client.agents.create_version has no root-level kind and a flat definition dict.
• Full suite: 808 passed, 1 skipped.

Follow-up

This needs a patch release (v0.3.2) so the workflow templates pick the fix up via PyPI on the next pip install agentops-toolkit on CI.