Add spreading activation, temporal recall, and retrieval quality improvements

[not-reed]

Summary

• Spreading activation for graph traversal — scored BFS with edge-weight-aware decay, replacing flat depth-only traversal. Scores never exceed parent (weights reduce decay, don't amplify). New spreadActivation, searchNodesWithScores, getRelatedMemoriesWithScores functions in cairn.
• Temporal filtering — since/before ISO date params on recallMemories and memory_recall tool, letting the LLM handle natural language date parsing.
• FTS5 stop word filtering — ~100 common English words stripped from FTS5 OR queries to reduce common-word noise in recall results.
• Retrieval quality across all apps: Construct graph memories now carry activation scores; Cortex signals stored with embeddings + graph extraction; Loom now uses graph expansion (was entirely missing).
• Observability — debug logging for injected skill instructions (with scores), relevant memories (with match type + score). Skill selection threshold raised 0.35 → 0.45.
• Generation priority — observation budget eviction now factors in generation counter (survived more compression rounds = evicted last).

Test plan

• 353 tests passing (14 new tests for spreading activation, temporal recall, graph score propagation, stop words)
• Typecheck clean across all packages
• Lint + format clean
• Manual testing via live Telegram bot — verified skill selection, memory recall, and logging output
• Monitor cairn VM after deploy for regression

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying sprawl with    Cloudflare Pages

Latest commit:	4bba4d9
Status:	✅  Deploy successful!
Preview URL:	https://ef02138b.sprawl.pages.dev
Branch Preview URL:	https://feat-memory-retrieval-improv.sprawl.pages.dev

View logs