Chronicle: fix sessions.summary storing raw JSON instead of human-readable text

[Copilot]

sessions.summary was being populated verbatim from USER_REQUEST span attributes or first user message events, which can be structured JSON (e.g. [{"type":"tool_result","content":"..."}]). This broke LIKE-based queries and confused the Chronicle intent LLM.

Changes

• New helper extractPlainTextFromContent (sessionStoreTracking.ts): detects JSON content (starts with [ or {) and extracts human-readable text from known message structures:

  • {"type":"text","text":"..."} — multi-modal text parts
  • {"role":"...","content":"..."} — chat messages with string content
  • Returns undefined for unrecognized JSON (tool results, images, etc.), skipping the summary entirely
  • Falls back to original string if JSON parsing fails

• _backfillFromSpanAttributes and _handleAgentSpan now route through the helper before setting summary, and skip the write when undefined is returned

• 13 new tests in sessionStoreTracking.spec.ts covering plain text pass-through, text part/chat message extraction, tool result skipping, empty/invalid JSON, and whitespace handling

// Before: stored verbatim
summary = '[{"type":"tool_result","content":"..."}]'

// After: skipped (no human-readable text extractable from tool_result JSON)
summary = undefined

// Text parts and chat messages are still extracted correctly
extractPlainTextFromContent('[{"type":"text","text":"Fix the login bug"}]')
// → "Fix the login bug"