Tiered search: narrow-to-wide retrieval strategy

[jordanpartridge]

Problem

Current search queries the entire knowledge base uniformly. This is wasteful — most queries are answered by recent, high-confidence entries. Searching everything equally burns tokens and can surface stale or low-relevance results.

Proposal

Implement tiered search that starts narrow and widens only if needed:

Search Tiers (in order)

1. Working context — Current session/project active entries
2. Recent entries — Last 14 days of daily logs
3. Structured storage — Permanent entries matching the query domain
4. Archive — Historical entries, only if earlier tiers yield nothing

Behavior

• Search starts at Tier 1 and returns early if confident matches are found
• Each tier adds latency/tokens, so early return is the optimization
• User can force a specific tier: know search --tier archive "old config"
• Results include tier source so the user knows where info came from
• Confidence and freshness factor into result ranking

Result Ranking

score = relevance * confidence_weight * freshness_decay

Where:

• relevance: semantic/keyword match score (Qdrant)
• confidence_weight: high=1.0, medium=0.7, low=0.4
• freshness_decay: exponential decay based on last_verified date

Benefits

• Faster results for common queries
• Token-efficient (don't load archive for simple lookups)
• Fresher, higher-confidence results surface first
• Historical search still available when needed

Acceptance Criteria

• Search implements tiered strategy with early return
• Results show which tier they came from
• --tier flag to target specific tiers
• Ranking incorporates confidence and freshness
• Archive search only triggered when earlier tiers insufficient

[jordanpartridge]

Shipped in PR #120 (tiered search with narrow-to-wide retrieval).