feat: hierarchical summaries — store verbatim + summary per warm row

[salishforge]

Summary

During LLM-mode consolidation, store both the full verbatim content AND a one-line summary per warm-tier row. Retrieval can return summaries for token-efficient context or drill into verbatim for precision.

Design

• Add summary TEXT column to warm_tier (nullable, populated only in summarize mode)
• During consolidation with LLM: summary = ConsolidationSummary.summary, content = raw verbatim
• query() returns both fields; client chooses which to use
• Budget-controlled retrieval (feat: budget-controlled retrieval — max_tokens parameter on query #64) prefers summaries when budget is tight

Inspiration

Combines FABLE (arXiv 2601.18116) multi-granularity concept with MemPalace (MIT) closets/drawers pattern — closets are summaries pointing to original content in drawers.

Migration

schema/migration-v2.5.sql: ALTER TABLE warm_tier ADD COLUMN IF NOT EXISTS summary TEXT;