Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

The google-vertex-anthropic provider uses AnthropicMessagesLanguageModel from the AI SDK, which supports Anthropic-style prompt caching via providerOptions.anthropic.cacheControl. The applyCaching() gate in transform.ts catches vertex-anthropic models implicitly via model.api.id.includes("anthropic"), but there's no explicit providerID check for google-vertex-anthropic. This is fragile if the api.id format changes.

For cache token tracking, the Anthropic SDK on Vertex sets response metadata under both the canonical "anthropic" key and a custom "vertex" key (derived from provider string vertex.anthropic.messages). The cacheWriteInputTokens extraction in session/index.ts only checks the "anthropic" key today.

Proposed changes:

1. Add model.providerID === "google-vertex-anthropic" to the applyCaching() gate condition for explicit, stable matching.
2. Add input.metadata?.["vertex"]?.["cacheCreationInputTokens"] to the cache write token extraction chain in session/index.ts as a defensive fallback.

Scope note: Native google-vertex (Gemini) is not included because Gemini uses implicit server-side caching, not Anthropic-style per-message cache breakpoints. Gemini's implicit caching already works without any client-side changes (verified: 97.8% cache hit on second request with identical prefix).

Verified behavior:

• Vertex Anthropic (Claude on Vertex): cache write tokens tracked correctly (46K tokens on a cold request).
• Vertex Gemini (gemini-2.5-flash): implicit caching verified end-to-end via test script. Second request with identical prefix showed cachedContentTokenCount: 28645 out of 29,293 input tokens, correctly normalized to cachedInputTokens by the AI SDK.
• Both providers work with bun run dev against live Vertex API endpoints.