Can I vibecode Wireflow?

Build a highly capable, production-ready local visual pipeline runner for AI media generation, inspired by node-based editors, from scratch in an empty repository.
Requirements and constraints must be followed strictly:

Architecture & Stack:
- Implement exactly one Vite + React frontend application and one Express server backend, both residing in the same monorepo.
- Use SQLite via `better-sqlite3` for local state persistence.
- Do NOT implement authentication, user accounts, or multi-tenant logic. The application must run solely on localhost.
- Provide a responsive Node Canvas built on `React Flow`. Users must be able to drag nodes from a side palette, wire outputs to inputs, pan, zoom, and have the graph autosave to SQLite on every single change.

Node Types & Execution:
- Implement exactly six node types and absolutely no more:
  1. Text input
  2. Image import (which uploads files to a local `./storage` directory)
  3. Prompt template (interpolates `{{upstream}}` values dynamically)
  4. Image generate
  5. Video generate
  6. Preview
- Generate nodes must call `fal.ai` exclusively through the `@fal-ai/client` library, expecting a `FAL_KEY` in the `.env` file.
- Model configurations must live in a `models.json` file. For each model, define its fal.ai path and explicitly map which fields are exposed as wired input ports versus which are static typed-in settings. Ship exactly two entries in this file: `fal-ai/flux/dev` and `fal-ai/kling-video/v1/standard/text-to-video`.

Execution Engine Rules (CRITICAL):
- When a run is triggered, perform a topological sort of the graph and execute it layer by layer.
- Store each node's output in SQLite, keyed by a composite of the node ID and a hash of its resolved inputs. Re-running the pipeline must only re-execute nodes whose input hash has changed (caching).
- Generate nodes must NEVER run automatically on edit. Only a global "Run" button or a specific per-node "Run" button can submit a job. Do not add reactive auto-execution anywhere in the application.
- Submit generation jobs exclusively through fal.ai's queue API and poll for completion by request ID. You MUST persist the request ID to SQLite BEFORE the first poll, ensuring that a pending job can survive and resume polling after a server restart.

Documentation & Exclusions:
- Create a comprehensive README that explicitly covers: where to place the `FAL_KEY`, a clear warning that every run spends real money at fal.ai, and detailed instructions on how to add a third model to `models.json`.
- Deliberately exclude: multiplayer collaboration, credits and billing systems, iterators or fan-out logic over arrays, complex video assembly/editing, hosted deployments, and integrations with any provider other than fal.ai.

Ensure the final product has clear error handling, loading states during generation, and sensible UI contrast.