v0.72.0

Deep search budget utilization, entity matching, inline edge backfill

• find(deep=True) triggers system doc migration and edge backfill inline so stores that never ran a write get edges on first search
• Entity matching uses word-boundary regex instead of tokenization, handling multi-word names and punctuation correctly
• Remove per-group deep item caps (5/group hard limit was wasting most of the token budget); renderer's budget now controls output volume
• Embedding search uses all edge source IDs instead of gating on FTS matches, so both signals contribute independently to RRF
• Reduce exclude set to match deep_primary_cap=3 so items at positions 4-9 aren't invisible
• Render order: deep items before detail backfill, so evidence gets budget priority and remaining tokens fill parts/versions