feat(fal): add 14 new FAL endpoints including Flux 3 Video Edit

[georgi]

What changed

Diffs the live FAL catalog (1500 models, fetched page by page from fal.ai/api/models) against the endpoints packages/fal-codegen/src/configs/ already carries, and adds every model published since 2026-08-26 that exposes a queue OpenAPI schema — 14 endpoints, headlined by Flux 3 Video Edit (blackforestlabs/flux-3/edit-video), which re-renders an existing video from a natural-language instruction and completes the flux-3 video family alongside the extend/draft/enhance nodes already shipped.

Module	Endpoints
video-to-video	blackforestlabs/flux-3/edit-video, google/gemini-omni-flash/v1.1/edit, fal-ai/kling-video/o3/4k/video-to-video/{reference,edit}, bria/video/background-removal/green-screen-despill
image-to-video	minimax/h3-max/multi-angle/image-to-video, minimax/h3-max/reference-to-video, google/gemini-omni-flash/v1.1/{image,reference}-to-video
text-to-video	google/gemini-omni-flash/v1.1/text-to-video
image-to-image	bria/increase-resolution
image-to-json	bria/ad-delayer, fal-ai/vggt-1b
speech-to-text	fal-ai/elevenlabs/forced-alignment

Manifest entries come from scripts/append-new-endpoints.ts, which fetches only what the manifest is missing — the 1586 existing entries are byte-identical (verified by comparing the parsed old manifest against the new one's prefix) and the 14 are appended. Class names follow the neighbouring precedents in each module (KlingVideoO34kVideoToVideoEdit after KlingVideoO34kImageToVideo, GeminiOmniFlashV11Edit after GeminiOmniFlashEdit, and so on); no collisions across 1600 configs.

Two candidates are deliberately left out: minimax/h3-max/director and fal-ai/controlfoley both return an empty queue OpenAPI document, so the generator has nothing to parse.

Unit pricing is not in this diff — api.fal.ai/v1/models/pricing needs a FAL_KEY this environment does not have. The new endpoints join the 166 already absent from the pricing bundle; estimateFalCost returns null for an unpriced node rather than guessing, and the next npm run generate:fal with a key picks all of them up.

Verification

• npm run test:affected — the 11 selected backend packages run; 69 of 72 turbo tasks pass, including @nodetool-ai/fal-codegen (152 tests) and @nodetool-ai/fal-nodes (161 tests). Two failures are pre-existing and reproduce on a clean tree (confirmed by stashing the diff):
  • @nodetool-ai/example-nextjs-vercel#build — Module not found: openclaw/fs-safe under packages/storage/node_modules
  • @nodetool-ai/base-nodes — 14 image tests failing with No WebGPU adapter available, the missing-Vulkan-driver case documented in AGENTS.md
• npm run typecheck — one pre-existing error in web/src/components/projects/__tests__/projectStarters.test.ts (isPersonal missing from a cast fixture); no file in this diff is involved
• npm run lint — passes; warnings only, all pre-existing
• npm run dev:nodetool -- harness gate --base origin/main — 4/4 selfchecks passed
• npm run generate:fal:check — 5 generated outputs match the checked-in fixtures, so the drift gate is still clean
• All 14 nodes register and resolve: FAL_NODES reports 1600 node types and every new type is present; nodetool node run fal.video_to_video.Flux3EditVideo reaches the provider and stops at FAL_API_KEY is not configured, which is as far as this environment goes

Agent capabilities

No capability added and no declared contract changed.

New checks

No check, rule, or audit added.

🤖 Generated with Claude Code

https://claude.ai/code/session_01R4Fb5v9vMJmS8xEpVp5Lyx

---

Generated by Claude Code