sdk: keep mid-conversation system messages positional

[aaliboyev]

Anthropic accepts role: "system" inside messages on current models, and some harnesses (Claude Code) now send it. The anthropic and gemini translators previously hoisted every canonical system/developer item into the top-level system field / systemInstruction, which

• invalidated the upstream prompt-cache prefix whenever a system message arrived mid-conversation, and
• could leave the messages array ending in (or containing back-to-back) assistant turns after removing the system item — an accidental prefill, rejected by models that no longer support prefilling.

Now only items leading the transcript merge into the start-of-conversation system form. Mid-conversation items stay positional, with canonical semantics following the OpenAI shape:

• openai (CC + Responses): system/developer items pass through in place natively, as before.
• anthropic + gemini: positional items are delivered in place as user turns wholly wrapped in <system> markers — works on every model (no capability matrix), keeps position, and leaves the cached prefix untouched. On anthropic they are deferred past an open tool round so a tool_use/tool_result pair is never split.
• hoist: true (new field on canonical message items): explicit opt-in to full system authority — the item merges into the top-level system / systemInstruction / leading system message on every shape, at the documented cost of invalidating the cached prefix. sdk/v1 owns the marker wrap/unwrap pair and the hoist splitter so the forms are identical across adapters.
• The marker round-trips: parse restores the system role on an exact whole-message match; substring occurrences are never promoted. Inbound anthropic parse now maps role:system messages to canonical system items positionally (previously they fell through to user).

Behavior documented in docs/reference/inference-api.mdx.