v0.52.1 — JSON mode kept half its promise on Claude

JSON mode kept half its promise on Claude. response_format: "json_object"
says "no markdown fences" and "works across all providers". On the native
claude-* path it injected the instruction and never stripped the fence Claude
adds anyway, so JSON.parse(response) threw on the leading backtick — on
exactly the models most likely to be asked for JSON.

Fixes

• fix(chat) — json_object unwraps a fenced reply on the native path.
  Only a reply that IS one fenced block is unwrapped; prose round a fence is
  the model ignoring the instruction and is returned as-is rather than trimmed
  into JSON that was never the whole answer. structuredContent.native keeps
  the fence as upstream evidence, and thinking blocks are untouched — the
  compat path would have done the strip for free at the cost of the native
  thinking channel. Pinned by six cases in
  test/chat-anthropic-json-fence.test.ts, including the verbatim
  haiku-4.5 reproduction. Reported by a sibling session reading the gateway's
  stripJsonFence call sites, neither of which is on /v1/messages.