feat: add fal.ai provider support

[kocaemre]

Summary

• Add fal.ai as an alternative generation provider while keeping Higgsfield as the default path
• Add fal.ai provider reference docs for scene stills, architecture-A legs, and architecture-B dives/connectors
• Add a small fal_pipeline.py helper that uploads local conditioning frames, submits fal jobs, writes JSON, and downloads returned media
• Update README/plugin metadata to mention fal.ai support

Test Plan

• python3 -m py_compile plugins/scroll-world/skills/scroll-world/references/fal_pipeline.py
• python3 plugins/scroll-world/skills/scroll-world/references/fal_pipeline.py --help
• python3 plugins/scroll-world/skills/scroll-world/references/fal_pipeline.py image --prompt test --out-json /tmp/nope.json exits with the expected missing-FAL_KEY error
• python3 -m json.tool plugins/scroll-world/.claude-plugin/plugin.json
• python3 -m json.tool .claude-plugin/marketplace.json
• git diff --check

Notes

• I did not run a live fal.ai generation because that requires a funded FAL_KEY; the adapter is documented around the published fal endpoint schemas and keeps field names configurable (--end-field tail_image_url for models that need it).

[oso95]

Closing as outdated — no reflection on the work, the ground moved underneath it.

This was written against v0.4.0. Since then the backend architecture was rewritten across v0.5.0–v0.8.0: Monid is now the default video-chain backend with a capability-check protocol (Step 0), providers are surfaced through the budget interview by measured cost, and the camera style became a user choice at intake. The PR conflicts in .claude-plugin/plugin.json, README.md, and skills/scroll-world/SKILL.md, and fal_pipeline.py predates the current pipeline contract — this is a rewrite rather than a rebase.

You also noted the adapter was never run live against a funded FAL_KEY, and the seam rule is the part of this project that has to be verified rather than inferred from endpoint schemas: a provider only qualifies once first/last-frame conditioning is probed and shown to frame-lock.

If you still want fal.ai support, a fresh PR against current main following the Step 0 capability-check pattern (as PR #12 does for OpenRouter) with a real probe run would be very welcome.