Investigate using bicep snapshot predictedResources for resource group ownership classification

[jongio]

Summary

Investigate whether bicep snapshot + predictedResources can supplement or simplify the resource group ownership classification pipeline introduced in #7603.

Context

In the PR #7603 discussion, @vhvb1989 identified that bicep snapshot's predictedResources output only includes resources the deployment creates — resources declared with the existing keyword are excluded. This could provide a fast, offline, deterministic signal for distinguishing created vs. pre-existing resource groups.

Investigation areas

1. Tier 0 / Tier 1 supplement: Can predictedResources serve as a "Tier 0" check that short-circuits the classification pipeline when Bicep source files are available?
2. Source file availability: azd down currently works purely from deployment metadata. Snapshot requires the original Bicep files — how often are they available at destroy time?
3. Adopted/redeployed RGs: Does snapshot correctly handle RGs that were created externally then deployed into (operation type would be Create but the RG pre-existed)?
4. Nested deployments and modules: Are RGs created via nested deployments or modules flattened into predictedResources?
5. Deployment outputs: RGs referenced via deployment outputs (e.g., output rgName string = resourceGroup.name) are not resources — how to handle these?
6. Deployment stacks path: The stacks path skips RG classification entirely — does snapshot interact with this?

Related

• PR: fix: prevent azd down from deleting pre-existing resource groups #7603 (4-tier classification pipeline)
• Original suggestion: fix: prevent azd down from deleting pre-existing resource groups #7603 (comment)