Add FableLoom visual canon bindings and capability-aware conditioning

[atomantic]

Part of #5377. Blocked by #5378.

Goal

Make FableLoom compile structured scene bindings into capability-aware visual conditioning so a canon-locked shot uses the approved character package and compatible local LoRAs instead of relying on prompt text alone.

Proposed approach

• Implement the structured entity bindings and typed reference roles from the visual-continuity plan.
• Resolve Universe characters, wardrobe, props, location state, temporal predecessor, approved identity assets, and all compatible local character LoRAs by stable ids.
• Declare backend capability manifests for reference roles, reference budgets, LoRA support, multi-character preservation, first/last frames, and extensions.
• Allocate inputs deterministically; record omissions and warnings rather than silently dropping required conditioning.
• Gate canon-locked video on an author-approved storyboard first frame.
• Keep text-to-video and unsupported combinations available only as clearly marked draft/degraded paths.

Acceptance criteria

• A canon-locked render records exact bindings, approved assets, adapter hashes/scales, compiler version, temporal source, and omitted inputs.
• FableLoom automatically considers existing compatible character LoRAs without persisting local LoRA paths in Universe records.
• Backends cannot claim support for conditioning roles they do not implement.
• A backend that cannot preserve the bound cast blocks canon-locked generation or requires an explicit draft override.
• Opening, convergence, and predecessor selection are deterministic and graph-aware.
• Provider/version compatibility and tests cover both supported and degraded paths.

Out of scope

• A second character registry.
• Assuming seeds alone lock identity.