fix(ai): drop empty unfinished reasoning items on replay

[rekram1-node]

Fixes #48319

Summary

When a Responses stream starts a reasoning item (response.output_item.added), OpenCode immediately persists session.reasoning.started with text: "" and state: { itemId: "rs_...", reasoningEncryptedContent: null } (encrypted_content only arrives on response.output_item.done).

If the server process dies or restarts abruptly mid-stream before session.reasoning.ended / session.step.failed is written:

1. to-llm-message.ts only checked message.error === undefined and therefore treated the uncompleted reasoning stub (time.completed === undefined) as having replayable providerMetadata.
2. OpenResponses.lowerMessages emitted an empty shell { type: "reasoning", id: "rs_A:rs_B", summary: [], encrypted_content: null } with neither summary text nor encrypted_content, which Responses backends reject with 400 Referenced reasoning item 'rs_A:rs_B' was not found or has expired on every subsequent turn.

Changes

• packages/core/src/session/runner/to-llm-message.ts: When an assistant message contains an uncompleted reasoning block (item.time !== undefined && item.time.completed === undefined), do not replay reasoning providerMetadata (state) and downgrade any non-empty partial summary text to "text", matching the existing message.error !== undefined path.
• packages/ai/src/protocols/open-responses.ts: After folding reasoning summary parts by itemId in lowerMessages, drop reasoning items that have neither summary text nor non-empty encrypted_content.