[Feature] Add Dreamverse mode selection for T2VA, FL2VA, and Ref2VA #1834

Motivation

Dreamverse currently starts every project through one implicit generation path. With FastH3 support now available, users need to declare whether a project is text-only, keyframe-conditioned, or driven by ordered multimodal references before generation begins.

This issue tracks the end-to-end mode-selection experience for the three MiniMax H3 task families:

• T2VA — text to video and audio.
• FL2VA — first/last-frame to video and audio.
• Ref2VA — ordered image, video, and audio references to video and audio.

Use these canonical identifiers in code and protocol payloads: t2va, fl2va, and ref2va.

Proposed scope

Frontend

• Add an accessible generation-mode selector to the pre-session Dreamverse composer.
• Default new projects to T2VA.
• Explain the input requirements for each mode.
• Keep the selected mode stable for a project and lock mode changes once its generation session starts.
• Show mode-specific inputs:
  • T2VA: prompt only.
  • FL2VA: first frame and optional last frame, subject to the final product decision.
  • Ref2VA: ordered image/video/audio references.
• Integrate selected assets from the planned Asset List / @ picker without duplicating asset ownership in the composer.

Protocol and runtime

• Include generation_mode and mode-specific conditioning inputs in project/session initialization.
• Validate mode names and conditioning requirements at the runtime boundary with actionable errors.
• Route T2VA/FL2VA requests to the H3 base pipeline and Ref2VA requests to the H3 reference pipeline/profile.
• Preserve the current WebSocket/session/GPU-pool ownership boundaries.

Tests

• Cover canonical mode metadata and the T2VA default.
• Cover selection, disabled/busy behavior, and project-init serialization.
• Cover runtime validation for valid modes, unknown modes, and missing/incompatible conditioning inputs.
• Add a mock-server path so the UI flow can be tested without GPUs.

Acceptance criteria

• A user can choose T2VA, FL2VA, or Ref2VA before starting a project.
• The UI clearly communicates and collects inputs required by the selected mode.
• The selected mode and ordered conditioning assets reach the runtime through a typed payload.
• The runtime rejects invalid combinations and dispatches valid requests to the correct H3 path.
• Existing LTX/Dreamverse flows remain backward-compatible when generation_mode is absent.
• Frontend, protocol, mock-server, and runtime tests pass.

Dependencies and coordination

• FastH3 Dreamverse backend foundation: Add H3 support into Dreamverse #1800 (merged).
• Asset List is tracked as Aryan's work in the Dreamverse V2 plan; this issue should consume that surface for FL2VA/Ref2VA attachment selection.
• Confirm whether FL2VA accepts first-only and last-only inputs in the product UI, or requires both endpoints.
• Confirm model/profile routing for FastH3 Preview versus base H3 checkpoints before enabling real FL2VA/Ref2VA generation.

Related resources

• Dreamverse V2 planning document: https://docs.google.com/document/d/1Dzty4rIXVNUXZDZH47ia6aw6Ad88RhgBcEMwJ43f2Fs/edit
• FastH3 Dreamverse integration: Add H3 support into Dreamverse #1800
• MiniMax H3: https://github.com/MiniMax-AI/MiniMax-H3