Lohnt sich dein eigenes Wireflow?

Build me a local visual pipeline runner for AI media generation. Requirements:

- One Vite + React app and one Express server in the same repo, SQLite via
  better-sqlite3, no auth, no accounts, runs on localhost.
- Canvas on React Flow: drag nodes from a palette, wire outputs to inputs,
  pan and zoom, autosave the graph to SQLite on every change.
- Six node types only: Text input, Image import (upload to ./storage), Prompt
  template (interpolates {{upstream}} values), Image generate, Video generate,
  Preview. Generate nodes call fal.ai through @fal-ai/client, FAL_KEY in .env.
- Model config lives in models.json: per model, its fal path and which fields
  are wired ports versus typed-in settings. Ship exactly two entries,
  fal-ai/flux/dev and fal-ai/kling-video/v1/standard/text-to-video.
- Run = topological sort, execute layer by layer, store each node's output in
  SQLite keyed by node id plus a hash of its resolved inputs. Re-running only
  re-executes nodes whose input hash changed.
- Generate nodes NEVER run on edit. Only the Run button or a per-node run
  button submits a job. This rule is the whole point, do not add reactive
  auto-execution anywhere.
- Submit through fal's queue API and poll by request id, persisting the request
  id BEFORE the first poll, so a finished job survives a server restart.
- README covers where FAL_KEY goes, that every run spends real money at
  fal.ai, and how to add a third model to models.json.

Excluded on purpose: multiplayer, credits and billing, iterators or fan-out
over arrays, video assembly, hosted deploys, and any provider besides fal.ai.