openai-chat: keep the client's cache breakpoints on system and tool_result

[CMGS]

Problem

A /v1/messages request served by an openai-chat model lost its prompt-cache breakpoints:

• system blocks were flattened into a string system message, which dropped their cache_control.
• each tool_result was flattened into a string role: tool message, which dropped its cache_control.

An agent loop through OpenRouter to a Claude model puts a breakpoint on the system prompt and another on the trailing tool_result, and it never read its cache. Turns that ended on a user text block, whose marker did pass through, paid the 1.25× cache-write price. Turns that ended on a tool_result read nothing. The billed total_tokens running above prompt + completion was that write premium, not an overcount: cost matches OpenRouter's reported cost.

Fix

• The system message now carries the native system blocks as content parts.
• A tool_result with cache_control becomes a one-part text array that keeps its marker. An unmarked one stays a string.
• Tool-definition markers are still dropped. Gemini's OpenAI-compatible endpoint rejects a cache_control field on a tool (400), and any later breakpoint already covers the tools prefix.

Evidence

• Gates (macOS): cargo fmt --check and cargo clippy --all-targets -- -D warnings are clean. cargo test passes 705, fails 0, ignores 4. The PG/Redis env-gated suites were not run because no storage code changed.

• Fail-first: native_system_blocks_stay_inside_the_openai_wire_shapes and tool_result_cache_control_survives_the_openai_wire both fail on the unfixed source.

• Live test: a release build sent requests through a recording proxy to OpenRouter, model anthropic/claude-sonnet-4.6. Each shape went out twice; the table shows the second call's cache read.

  breakpoint on	before	after
  system blocks	0	3227
  user text block	3238	3232
  trailing tool_result (+ system)	0	3879

  In a 4-turn growing agent loop, cache reads are 0 / 3885 / 3968 / 4051 tokens after the fix. On every ledger row, cost_micros is within 1 micro of the vendor-reported cost.

• Vendor acceptance (direct calls): a system content-part array (with and without cache_control) and a tool message content array with cache_control both return 200 on OpenAI, DeepSeek, Qwen (compatible mode), Moonshot, xAI, MiniMax, OpenRouter and Gemini's OpenAI-compatible endpoint.