Backport: feat(google): support agentic and static video processing on generateContent

[http-samc]

This is a backport of #21925 to the release-v6.0 branch. It was ported manually because the automated cherry-pick conflicts (see below). FYI @http-samc

Background

Same as #21925. Gemini agentic video understanding (mediaProcessing) and static clipping/frame rate (videoMetadata) were unreachable from google(...) / generateContent, and the option was silently dropped.

Summary

This is a port of #21925 to the @ai-sdk/google 3.0.x layout. It changes convert-to-google-generative-ai-messages.ts and google-generative-ai-prompt.ts, and resolves namespaces via getGoogleProviderOptions.

• providerOptions.google.processing on user video/* file parts maps as follows:
  • 'agentic' becomes mediaProcessing: 'AGENTIC'.
  • 'static' becomes mediaProcessing: 'STATIC'.
  • { type: 'static', startOffset?, endOffset?, fps? } becomes STATIC plus videoMetadata.
  • Invalid values produce a warning and are dropped.
• Docs are in 15-google-generative-ai.mdx. Unlike on main, there is no GoogleInteractionsVideoOptions type and no Interactions cross-link, because this branch's Interactions provider doesn't support agentic video.
• Example: examples/ai-functions/src/generate-text/google/agentic-video.ts.
• Patch changeset.

Why manual: git cherry-pick of #21925 onto release-v6.0 conflicts on convert-to-google-generative-ai-messages.ts, google-prompt.ts (absent here) and the docs file. This branch is named backport-pr-21925-to-release-v6.0, so if the backport label is applied to #21925 the workflow detects this PR and skips instead of opening a conflict-marker PR.

End-to-End Verification

Same live check as #21925, run against this branch's source: gemini-3.7-flash with a 10s clip that has 8 flashes in total (4 in the first 5s).

processing	release-v6.0 today	this branch
(unset)	"2", 646 input tokens	"2", 646 input tokens
'agentic'	"2", 646 tokens (ignored)	"8", 108 input tokens*
{ type: 'static', startOffset: 0, endOffset: 5, fps: 5 }	"2", 646 tokens (ignored)	"4", 1,591 input tokens

* On this branch, usage.inputTokens excludes toolUsePromptTokenCount (Google reported about 35k tool-use tokens for this request). The fix is #21033 on main, which cherry-picks cleanly onto release-v6.0. Please apply the backport label to #21033 so the workflow carries it over.

Package checks: vitest (node: 647 passed; edge for the changed suites: 236 passed), tsc --build, oxfmt, oxlint.

Checklist

• All commits are signed (PRs with unsigned commits cannot be merged)
• Tests have been added / updated (for bug fixes / features)
• Documentation has been added / updated (for bug fixes / features)
• A patch changeset for relevant packages has been added (for bug fixes / features - run pnpm changeset in the project root)
• I have reviewed this pull request (self-review)

Related Issues

Backport of #21925. Continues to release-v5.0 in #21927.