fix: prefer mental model abstraction in reflect to avoid unnecessary low-level retrieval

[Oxygen56]

Summary

The reflect operation was unconditionally forcing the full hierarchical retrieval chain (search_mental_models -> search_observations -> recall) before allowing the agent to answer, even when mental model abstraction would provide adequate context. This PR adds budget-aware short-circuit logic to prefer higher-level abstractions when appropriate.

Problem

In run_reflect_agent, the forced_sequence always forced search_mental_models, then search_observations, then recall before allowing auto mode. Even when a fresh, relevant mental model could answer the query, the agent was forced to continue to lower-level retrieval, increasing latency, cost, and potentially duplicating already-synthesized knowledge.

Changes

1. Budget-aware short-circuit in forced sequence (agent.py lines 599-618):

   • low budget: Force only search_mental_models, then allow auto mode
   • mid budget: Skip lower-level retrieval when mental models are fresh and non-empty
   • high budget: Preserve the full hierarchical verification path (existing behavior)

2. Mental model freshness tracking (agent.py line 425):

   • Added mental_models_sufficient flag to track when retrieved mental models are sufficient

3. Freshness evaluation (agent.py lines 982-990):

   • After search_mental_models returns, check for mental models with is_stale=False and non-empty content
   • Set mental_models_sufficient = True when at least one model meets these criteria

Behavior

Budget	Before	After
low	Force all 3 retrieval layers	Force only search_mental_models, then auto
mid	Force all 3 retrieval layers	Force search_mental_models; skip lower layers if fresh results found
high	Force all 3 retrieval layers	Force all 3 retrieval layers (unchanged)

The default budget (low) now completes in fewer iterations when mental models are available, reducing latency and cost while maintaining answer quality.

Backward Compatibility

• high budget: Unchanged behavior (full verification chain)
• mid budget: Conditional short-circuit only when mental models are truly fresh
• low budget: Changed default to prefer speed, consistent with documented budget semantics ("Prioritize speed over completeness")

Fixes #1971