feature: SDK should only upsert fields that have been modified

[minitriga]

Component

Python SDK

Describe the Feature Request

When saving a node (especially with allow_upsert=True), the SDK currently includes all fields in the GraphQL mutation payload, not just fields that have been modified. This can lead to unintended side effects, such as:

• Sending null for unset optional relationships (e.g. parent: null for hierarchical nodes), causing server-side errors
• Overwriting values that were set by other processes, object templates, or server defaults
• Unnecessary data in mutation payloads increasing request size and processing time

The SDK should implement dirty/change tracking so that _generate_input_data() only includes fields that have been explicitly set or modified since the node was created or fetched.

PR #810 addressed a specific instance of this problem (hierarchical relationships), but the underlying issue applies more broadly to all attributes and relationships.

Describe the Use Case

1. Upsert without overwriting: A user fetches a node, modifies one attribute, and saves it. Currently all fields are sent, potentially overwriting concurrent changes to other fields. Only the modified attribute should be included in the mutation.

2. Object template defaults: When creating a new node with allow_upsert=True, unset optional fields are sent as null, which can override defaults that would otherwise be applied by object templates on the server side.

3. Concurrent workflows: Multiple scripts or users may modify different fields on the same node. Without change tracking, each save overwrites the entire object, creating race conditions.

Additional Information

• Related PR: Fix: exclude uninitialized hierarchy relationships from mutation data #810 (fixes the specific case of hierarchical relationships)
• The fix in PR Fix: exclude uninitialized hierarchy relationships from mutation data #810 skips hierarchy relationships unless explicitly initialized — this pattern should be generalized to all fields
• Key area of the codebase: infrahub_sdk/node/node.py — _generate_input_data() method
• A possible approach is to track a _dirty set on each node that records which attributes/relationships have been explicitly set or modified, and filter _generate_input_data() output accordingly