Fix period-based content splitting in external paper search cache key

[tnkshuuhei]

Problem

buildStableKey() and parseContent() in lib/external-paper-search.ts split content on the first period (.) to separate title from description:

function buildStableKey(fromContent: string, toContent: string): string {
  const fromTitle = fromContent.split(".")[0].trim();
  const toTitle = toContent.split(".")[0].trim();
  return `${fromTitle} → ${toTitle}`.trim();
}

Card titles containing periods (e.g., "Dr. Smith's Intervention", "e.g. community health programs", "U.S. education policy") are truncated incorrectly:

• "Dr. Smith's Intervention" → "Dr"
• "e.g. community health" → "e"

This affects:

1. Cache keys: Truncated keys may collide or miss, though worst case is a cache miss (not incorrect results)
2. LLM query extraction input: parseContent() feeds truncated titles to Gemini 2.5 Flash, potentially degrading search query quality

Proposed Solution

• Use a more robust separator (e.g., the first sentence-ending period followed by a space and uppercase letter)
• Or pass title and description as separate fields from the caller instead of concatenating with period separator
• Or hash the full content for cache keys

Relevant Files

• lib/external-paper-search.ts:60-75 — buildStableKey() and parseContent()
• mastra/workflows/logic-model-with-evidence.ts:326-331 — Caller that concatenates title. description

Context

Identified during review of PR #194.