Cache surplus practice blocks for instant delivery

[sfw]

Summary

• When the LLM generates 2-3 practice_problem blocks, only the first is delivered — extras are cached as separate GeneratedContent entries via SurplusPracticeCache
• On the next continue request, cached surplus blocks are served instantly before falling back to LLM generation, eliminating wait time
• Surplus entries use is_predictive_warm: true to piggyback on existing mastery-change invalidation — no changes to invalidation infrastructure needed

Test plan

• 12 new unit tests covering split logic, cache/pop ordering, session isolation, invalidation flags, and TTL
• Backend tests: 733 passed
• Frontend tests: 289 passed
• Backend lint clean
• Frontend lint clean
• Frontend production build passes
• Manual: verify only 1 practice question appears per page
• Manual: verify next question loads instantly from surplus cache

🤖 Generated with Claude Code