Preserve sibling tool calls in assistant imports

[colinbennettbrain]

Summary

Fixes Lingua import handling for assistant role messages that contain Anthropic-style content[].type = "thinking" blocks alongside sibling OpenAI-style tool_calls.

The importer now parses this mixed assistant shape at a typed compatibility boundary and preserves both the reasoning content and sibling tool calls in the resulting universal assistant message.

It also avoids preempting OpenAI/OpenRouter/vLLM-style messages with top-level reasoning or reasoning_signature, allowing the existing OpenAI provider parser to preserve those fields.

Root cause

try_parse_mixed_role_messages_for_import offered role-message items to provider parsers before the lenient fallback. A provider parser could preserve content[].thinking while ignoring sibling tool_calls. The lenient assistant fallback only parsed content, so sibling tool calls were invisible there too.

Tests

• CASE_FILTER=role-message-thinking-with-sibling-tool-calls cargo +1.95.0 test -p lingua --test import_fixtures -- --nocapture
• CASE_FILTER=role-message-openai-reasoning-with-sibling-tool-calls cargo +1.95.0 test -p lingua --test import_fixtures -- --nocapture
• cargo +1.95.0 test -p lingua --test import_fixtures -- --nocapture
• RUSTUP_TOOLCHAIN=1.95.0 make typed-boundary-check