fix(openai): replay reasoning text signatures

[arjunkmrm]

Responses-compatible gateways that serve Anthropic models return thinking as reasoning_text content with a signature on the reasoning item, and leave encrypted_content empty. The provider decoded away the signature and never replayed reasoning content, so with store: false the next request carried an empty reasoning item and the model lost its prior thinking every turn.

• OpenAiSchema: reasoning_text content accepts an optional signature.
• OpenAiLanguageModel: reasoning content is kept as reasoningContent metadata on non-streamed reasoning parts and on streamed reasoning-end (from output_item.done), and replayed as content on the reasoning input item. Items without content are unchanged.
• @tardie/ai-openai-compat: declares the same reasoningContent field, since both providers augment the shared Prompt and Response metadata types and the declarations must match.
• CLAVIA_PATCHES.md: new "Reasoning text signatures" row.
• Two tests: non-streamed replay sends the signed content back; streamed reasoning-end and Prompt.fromResponseParts keep it. Both fail without the change. Release workflow steps pass locally: lint, check, provider tests (377), build, pack and the packed consumer check.