feat(skills): add Atlas Cloud media generation

[binyangzhu000-sudo]

Summary

• Add a bundled Atlas Cloud media skill for executable image and video generation.
• Submit asynchronous predictions, use bounded polling, validate HTTPS output URLs, and download media without forwarding API credentials.
• Provide model overrides, model-specific parameters, safe 64 MiB downloads, documentation, and offline contract tests.

Related issues

None.

Type of change

• fix — bug fix
• feat — new feature
• docs — documentation only
• refactor / perf / test / chore / ci

Checklist

• npm run typecheck passes
• npm test passes
• npm run build compiles
• Added/updated tests for the change
• No secrets, real personal data, or local absolute paths in the diff
• PR title follows Conventional Commits

Notes for reviewers

• The focused skill validation and 4 Python contract tests pass.
• Real Atlas Cloud calls generated a 1024x1024 PNG with qwen-image-3.0/text-to-image and a valid MP4 with bytedance/seedance-2.0-fast/text-to-video; generated media is not committed.
• The full test command emitted 998 passing assertions and no not ok result, but did not exit after more than seven minutes because the suite retained a Node handle, so the full-suite checkbox remains unchecked.
• The implementation uses only the Python standard library and does not add runtime dependencies.

[Bhasvanth-Dev9380]

Hey, thanks a lot for this! I went through it properly and the Atlas Cloud
integration is solid. I checked the endpoints, the model IDs, the response
shape and the size format against their docs and you got all of it right.

One request before we merge: would you be up for rewriting it in TypeScript?

The reason is just that Brigade pins Node >= 22.12, so Node is there on every
install, but Python isn't guaranteed. If someone doesn't have python3, the skill
quietly disappears from the agent's list and they never find out why. And where
python3 does exist it's often 3.9 (macOS system python), which is where your
timeout handling has a small gap.

There's already a node example in the repo if it helps:
skills/sherpa-onnx-tts/bin/sherpa-onnx-tts.

The other nice part is that a TS version lifts almost straight into
generate-video-tool.ts as a proper provider later — same fetch/Buffer idioms —
so the work would count twice instead of being rewritten.

Two small things worth carrying over while you're in there:

• The download only checks Content-Length against the 64 MiB cap, not against
  bytes actually received, so a truncated download currently saves a partial
  file and reports success.
• The poll ceiling works out to 5 minutes (150 x 2s), but your SKILL.md says
  video can take several minutes — ours is 12. Also worth printing the
  prediction id on timeout so a paid generation isn't lost.

Happy to merge once it's in TS. And if you have any thoughts on this, or you
think I'm wrong somewhere, please do get back — I'd rather hear it than have you
just go along with it. Really appreciate the effort you put in here.