Track `ChatAnthropicVertex` prompt-caching parity and provider consolidation

Requested by @sydney-runkle. Tracker for the missing automatic prompt-cache setup when Deep Agents is given langchain_google_vertexai.model_garden.ChatAnthropicVertex. This does not propose closing either linked PR.

Observable gap

Deep Agents currently appends its prompt-cache stack to the main agent, the automatic general-purpose subagent, and declarative inline subagents. Its always-present AnthropicPromptCachingMiddleware only accepts isinstance(request.model, ChatAnthropic). ChatAnthropicVertex is a distinct BaseChatModel, so the middleware is invoked with unsupported_model_behavior="ignore" and silently passes the request unchanged.

Consequently, an Anthropic Claude model reached through Vertex does not receive the automatic system/tool annotations or model_settings["cache_control"] that direct ChatAnthropic receives. The immediate effect is missing automatic cache-control payloads, which can increase repeated-prefix input processing cost. It is not a promise that every request would otherwise be a cache hit: provider eligibility, minimum-size rules, a stable matching prefix, and cache lifetime still decide that.

The adapter path is ready: langchain-google #1631 is merged and has ChatAnthropicVertex._format_params() remove cache_control from top-level parameters and add it to the final formatted message content block, for sync, async, and streaming paths. That makes the following a temporary opt-in workaround for callers able to pass the model option (or an equivalent custom wrap_model_call middleware that sets model_settings["cache_control"]):

cache_control = {"type": "ephemeral", "ttl": "5m"}
model.invoke(messages, cache_control=cache_control)

Use a released langchain-google-vertexai containing #1631. This workaround targets the trailing message breakpoint; it is not a replacement for Deep Agents' full automatic system/tool handling.

Public-source audit

No public JS Vertex or Fireworks prompt-caching middleware was found in the current langchainjs middleware tree. Gemini is intentionally not treated as the same adapter issue: Mason's consolidation proposal notes its out-of-band CachedContent shape, whereas OpenAI prefix caching is automatic.

Requested tracking / acceptance criteria

Link this as a provider follow-up to #3239 — Generalize caching, the open consolidation issue authored by @mdrxy (Mason Daugherty). Its stated direction is a neutral PromptCachingMiddleware with provider dispatch, deprecated Anthropic alias, explicit supported-provider documentation, and no semantic change on the initial Anthropic path. This issue supplies the ChatAnthropicVertex case rather than broadening #3239 without review.

A complete solution should:

• integrate the Vertex handler through the neutral/provider-dispatch design rather than add another permanent ad-hoc gate;
• preserve existing Anthropic, Bedrock, and Fireworks behavior and make support/no-op behavior explicit;
• apply the same behavior to the main agent, automatic general-purpose subagent, and declarative inline subagents;
• cover synchronous and asynchronous calls, including the adapter's actual request payload and returned usage metadata where available;
• retain optional imports: Deep Agents must not gain hard dependencies on AWS, Fireworks, or Vertex integrations, and an unavailable or older optional integration must no-op without masking unrelated import errors;
• test a real formatted Vertex request for a content-block cache_control marker (not only a middleware ModelRequest), plus an unsupported-model no-op; and
• document that automatic markers optimize eligible matching prefixes, not guaranteed provider cache hits.

Related public work: langchain-google #1993 and deepagents #6318, both open at filing. Hunter Lovell (@hntrl) authored those PRs.