fix: include accumulated reasoning_details in reasoning-end event + regression test

[robert-j-y]

Description

Incorporates the fix from PR #394 (by @kalaomer) with formatting fixes, a patch changeset, and an e2e/issues/ regression test.

Bug: When streaming a text-only response (no tool calls) with reasoning enabled, the reasoning-end stream event was emitted without providerMetadata. The Anthropic signature arrives in the last reasoning delta, but reasoning-end didn't carry it forward. The AI SDK uses reasoning-end's providerMetadata to update the reasoning part, so the signature was lost — causing "Invalid signature in thinking block" on subsequent turns.

This was not an issue for tool-call responses, where accumulatedReasoningDetails were already attached to the first tool-call event.

Fix: Attach accumulatedReasoningDetails to reasoning-end's providerMetadata in both emission points:

1. When text content starts (reasoning ends before text begins)
2. In flush (reasoning ends when stream completes)

Before:

// reasoning-end emitted without providerMetadata → signature lost
controller.enqueue({ type: 'reasoning-end', id: reasoningId || generateId() });

After:

// reasoning-end now carries accumulated reasoning_details (including signature)
controller.enqueue({
  type: 'reasoning-end',
  id: reasoningId || generateId(),
  providerMetadata: accumulatedReasoningDetails.length > 0
    ? { openrouter: { reasoning_details: accumulatedReasoningDetails } }
    : undefined,
});

Important review notes

• Core fix authored by @kalaomer in fix: include accumulated reasoning_details in reasoning-end event #394 — this branch adds formatting fixes, changeset, and regression test on top
• E2e test was NOT run against live API — only unit tests were verified locally (210/210 pass on both node + edge). The e2e test targets anthropic/claude-sonnet-4 and will need a live run to confirm
• Second e2e test silently passes if content is a string — the if (typeof content === 'string') return; guard at line 106 means the multi-turn test won't assert anything if the AI SDK returns string content instead of structured parts

Checklist

• I have run pnpm stylecheck and pnpm typecheck
• I have run pnpm test and all tests pass
• I have added tests for my changes (if applicable)
• I have updated documentation (if applicable) — N/A, no doc changes needed

Changeset

• I have run pnpm changeset to create a changeset file

---

Link to Devin run | Requested by @robert-j-y