feat(gateway): take official hosts, output fallback and error shapes from core v0.42.0 - #42

Adopts the shared facts core v0.42.0 now carries, in place of this side's own copies.

Changes

• Official hosts. Upstream::is_official uses tw_dialect::official::is_official_host for standard upstreams (Azure and Bedrock stay official by shape). The local list had three hosts; core's also covers DeepSeek, Moonshot, Zhipu/Z.ai, DashScope, xAI, and *.amazonaws.com, and compares the parsed host, so https://relay/api.openai.com or https://api.openai.com@relay no longer passes as official.
• anthropic-version comes from tw_dialect::official::ANTHROPIC_VERSION.
• Default output length. When an upstream requires max_tokens and the caller set none, it is fallback_max_output_tokens(upstream model) (32000 for Claude, 8192 otherwise) instead of a fixed 4096 that cut Claude answers short. There is no per-model output limit stored here to prefer, so this mirrors the desktop's fallback branch.
• Stream error frames come from tw_dialect::convert::error_frame(client, status, message). Before, a Responses client got a Chat-shaped event: error frame it skips (the stream just stopped); it now gets response.failed. Anthropic clients get an error event whose error.type follows the status.
• Whole error bodies come from tw_dialect::convert::error_body in the caller's format. GatewayErrorResponse carries the client dialect; every generation handler sets it. Anthropic clients now get {"type":"error","error":{…}}. The Chat/Responses body's type is OpenAI's own vocabulary (rate_limit_error, server_error, invalid_request_error, …) instead of this gateway's tags (rate_limited, provider_http_error, …); status codes and Retry-After are unchanged.
• The provider connection test builds its URL with tw_dialect::url::upstream_url. tw_dialect::usage was already the only usage parser; tokens() uses Usage::prompt_total.

Tests

• Unit: official-host spoofing, error body per dialect.
• Integration (gateway_error_shapes.rs): Anthropic body on an upstream failure and on a refusal before routing, Chat error type, a failing Responses stream ending in response.failed (with the caller's model), a failing Anthropic stream ending in an error event.
• Local: fmt, clippy (--all-targets and --lib), 679 unit tests, full integration suite with own containers. The only failures were three tests that resolve public hosts (AWS docs MCP, two MCP tool tests hitting "URL host could not be resolved"), which depend on this machine's DNS and are untouched here.

🤖 Generated with Claude Code