[bug]: v7 Canvas & Control Layer metadata is not recorded or recallable (plus missing v4–v6 metadata recall fallbacks) #10131

Is there an existing issue for this problem?

• I have searched the existing issues

Version number

v7.0.0-alpha.4-post1 (main)

What happened

In the move from InvokeAI v6 (webv1) to v7 (webv2), two related metadata regressions were introduced for Canvas generations and parameter recall:

1. Canvas & Control Layer metadata is no longer recorded on generated images when generating on the v7 Canvas (ControlNets, T2I-Adapters, ControlLoRAs, Fun-ControlNet-Union, ControlNet-LLLite, mask-based Regional Guidance, and Canvas layer state).
2. Frontend parameter recall (imageRecall.ts) no longer displays or recalls Control / Canvas metadata and dropped fallback parsers for historic v4, v5, and v6 metadata schemas, meaning users cannot inspect or recall adapters and settings from either newly generated v7 Canvas images or their existing v4–v6 image galleries.

1. What Is Missing in v7 Metadata Recording (compileCanvasGraph.ts & CoreMetadataInvocation)

1. Control Layers (ControlNet, T2I-Adapter, ControlLoRA, Fun-ControlNet-Union, ControlNet-LLLite) are not recorded in core_metadata:
   • CoreMetadataInvocation (invokeai/app/invocations/metadata.py) no longer has inputs for controlnets, t2iAdapters, or Canvas control layers (only anima_inpaint_adapter / anima_inpaint_adapter_weight for Anima's inpaint adapter).
   • compileCanvasGraph.ts and addControlLayers.ts (invokeai/frontend/webv2/src/features/generation/core/canvas/) wire the control adapter nodes into denoise_latents, but write nothing to core_metadata.
   • Furthermore, graph.ts wires core_metadata directly into canvas_output.metadata (bypassing denoise_latents). Even with PR feat(nodes): add native metadata input and output across denoise invocations #10083 adding native controlnets and t2iAdapters serialization on all denoise nodes (denoise_latents.metadata), Canvas/Generate graphs overwrite canvas_output.metadata with core_metadata.
2. Mask-Based Regional Guidance is not recorded:
   • While layout-prompt models (ideogram-4, moss-image, flux3) record their assembled layout caption in core_metadata (ideogram4_caption, moss_image_layout_prompt, flux3_caption), mask-based Regional Guidance compiled by addRegionalGuidance.ts for SD1.5, SDXL, FLUX, Z-Image, Anima, and Krea-2 records none of the regional prompts, masks, or regional reference images in core_metadata.
3. Canvas Layer State (controlLayers, regionalGuidance, inpaintMasks, rasterLayers):
   • In v6 (webv1/src/features/controlLayers/store/selectors.ts), selectCanvasMetadata recorded canvas_v2_metadata (controlLayers, inpaintMasks, rasterLayers, regionalGuidance) on generated images, and CanvasLayers (webv1/src/features/metadata/parsing.tsx) displayed the layer count and recalled them via canvasMetadataRecalled. In v7, none of this Canvas layer state is recorded or recallable.

2. Missing Frontend Recall & Historic Metadata Format Support (v4 / v5 / v6 -> v7)

Users have large existing image libraries generated in InvokeAI v4, v5, and v6 (as well as node workflows). While backend workflow nodes (metadata_to_ip_adapters, metadata_to_controlnets, metadata_to_t2i_adapters in PR #10083) support multi-version recall across v4, v5/v6, and v7 schemas, the v7 frontend (invokeai/frontend/webv2/src/workbench/image-actions/imageRecall.ts) lacks both UI recall for Control/Canvas layers and fallback parsers for historic metadata formats:

What you expected to happen

1. Scope & Expectations — Staging Area / Direct Generation vs. Flattened Canvas Export:
   • In Scope (Staging Area & Direct-to-Gallery Generations): Metadata recording and recall should be complete whenever an image is produced by a generation run — i.e., when generating with destination set to Gallery, or when saving an individual generated candidate from the Canvas Staging Area to the Gallery. Each staged candidate is the direct output of a single compileCanvasGraph execution with well-defined model, prompt, seed, strength, Control Layers, Reference Images, and Regional Guidance parameters.
   • Out of Scope (Saving the Flattened Canvas Document on Its Own): Saving/exporting the composited Canvas raster layers outside the staging area is a document export that may combine manual brush edits, imported assets, and multiple separate generation passes, so it is not expected to carry single-generation metadata.
2. Record Control & Canvas Metadata in v7:
   • Record active Control Layers (ControlNet, T2I-Adapter, ControlLoRA, Fun-ControlNet-Union, ControlNet-LLLite) and Regional Guidance in core_metadata when compiling Canvas graphs (or route core_metadata through denoise_latents.metadata so denoise-emitted controlnets/t2iAdapters are preserved on canvas_output).
3. Display & Recall Control / Canvas Metadata in v7 UI:
   • Display used Control Adapters, Regional Guidance, and strength in the v7 Metadata Viewer.
   • Support recalling Control Layers, Regional Guidance, and strength back into the Canvas / Generate UI.
4. Preserve Backward-Compatible Recall for v4, v5, and v6 Images:
   • Add fallback parsers in imageRecall.ts for:
     • canvas_v2_metadata.referenceImages.entities and v4 ipAdapters -> v7 referenceImages
     • canvas_v2_metadata.controlLayers and v4/v7 controlnets / t2iAdapters -> v7 Canvas Control Layers
     • canvas_v2_metadata.regionalGuidance -> v7 Canvas Regional Guidance layers

Additional context

• Related PR: feat(nodes): add native metadata input and output across denoise invocations #10083 (Stage 1: Unify metadata across all 19 denoise invocations, including controlnets, t2iAdapters, and v7 ref_images serialization on denoise nodes and v4/v5/v6/v7 recall in metadata_to_ip_adapters, metadata_to_controlnets, and metadata_to_t2i_adapters).
• Related Discord #developer-forum discussion: https://discord.com/channels/1020123559063990373/1564749303069544558