Add concurrency limiter for Semantic Scholar API requests

[tnkshuuhei]

Problem

When external paper search is enabled, each under-matched edge triggers 1 LLM call (Gemini 2.5 Flash for query extraction) + 2 Semantic Scholar API requests in parallel. With Promise.allSettled in the workflow step, all edges are searched concurrently.

For a logic model with 15 under-matched edges, this fires 15 LLM calls + 30 Semantic Scholar API requests simultaneously. Semantic Scholar's free tier allows ~100 requests per 5 minutes, so larger logic models can easily hit 429 rate limit errors.

Because Promise.allSettled swallows these errors gracefully, the user experience degrades silently — the toggle is enabled but most edges return 0 external papers.

Proposed Solution

Add a concurrency limiter to searchExternalPapersStep in mastra/workflows/logic-model-with-evidence.ts or within lib/external-paper-search.ts.

Options:

• Batch processing: Process N edges at a time (e.g., 3 concurrent edge searches)
• Library-based: Use p-limit or similar to cap concurrent requests
• Built-in backoff: Add retry with exponential backoff on 429 responses in lib/academic-apis/semantic-scholar.ts

Relevant Files

• mastra/workflows/logic-model-with-evidence.ts — Step 2.5 (searchExternalPapersStep)
• lib/external-paper-search.ts — searchExternalPapersForEdge()
• lib/academic-apis/semantic-scholar.ts — executeSearch()

Context

Identified during review of PR #194. Not blocking for merge since the feature is opt-in (requires both NEXT_PUBLIC_EXTERNAL_SEARCH_ENABLED=true and user toggle), but should be addressed before wider adoption.