feat: chat() front door and Claude cache marking on LiteLLM routes

[rejojer]

Ports the eight pending commits from feat/local-chat (the v0.2.10 line) to main, following the #402/#404 pattern. After this merge, git diff main feat/local-chat is empty — main and the 0.2.10 branch are content-identical again (verified on the branch before opening this PR).

client.chat() — the answer-out front door

chat(messages, doc_id=None, stream=False, model=None) returns the answer string (a text-chunk iterator when streaming). It is thin sugar over chat_completions() in both modes — same wire, envelope unwrapped, nothing rewritten. Multi-turn is the caller-maintained role/content list; a bare query string works as the minimal call. This is the only format-unpinned contract: chat_completions() / responses() / messages() stay the protocol doors for callers who need the envelope or the process.

Managed-prefix cache marking on LiteLLM-routed Claude models

Claude prompt caching is opt-in per request (unlike OpenAI's automatic caching), so without a marker every managed turn re-pays full input price. chat_completions() now cache-marks the managed prompt prefix automatically on Claude models routed through LiteLLM — Anthropic direct, Bedrock Claude, Vertex Claude — via two documented parameter surfaces end to end: ModelSettings.extra_args → LiteLLM's cache_control_injection_points. No subclassing, no hooks, no content rewriting; user messages and the response envelope are untouched. Provider resolution asks LiteLLM's own get_llm_provider instead of hand-parsed prefixes.

Enablement doctrine: a wrong marker costs users a silent 1.25× cache-write premium, so each channel was live-verified write→read before being opened — Anthropic full-stack per-turn reads, Bedrock us.anthropic.claude-sonnet-5 (w7264→r7264), Vertex claude-sonnet-4-5 (w4842→r4842) — and non-Claude routes are pinned off by negative tests.

Also in the port

Review remediation and docs: the missing-openai-agents error now points anthropic-extra users at messages(); two silent vendor chains gained guards (recorder attachment asserted against the real OpenAIResponsesModel, and a wire-level test asserting cache_control reaches the Anthropic HTTP body through LitellmModel); the local text-only history contract is stated on chat_completions() (cloud forwards tool turns, local rejects; extra message fields drop); the max-tokens docstring trimmed to the contract.

Verification

276 tests green (3 skipped without the claude extra) on the ported branch. Cache marking was live-verified on all three Claude channels before landing on feat/local-chat, where #400 carries the full line for review.