Summary

• expose limit for search and limit/read_top_k for deep_search in the agent tool schema
• widen the default deep_search evidence pool from 5 to 10 while capping runtime requests at limit <= 20 and read_top_k <= 5
• add targeted DeepSeek eval notes for prior high-call misses

Validation

• uv run ruff format --check src/synaptic/agent_loop.py src/synaptic/agent_tools_v2.py tests/test_agent_efficiency.py tests/test_agent_tools.py
• uv run ruff check src/synaptic/agent_loop.py src/synaptic/agent_tools_v2.py tests/test_agent_efficiency.py tests/test_agent_tools.py
• uv run pytest tests/test_agent_tools.py tests/test_agent_efficiency.py tests/test_agent_search_benchmarks.py
• DeepSeek preflight: deepseek-v4-flash OK
• Targeted DeepSeek high-call miss eval: prior 0/6 reached -> new 2/6 reached (293992, 14151)