Implement tool selector (per-turn context shaping)

[martinjms]

Summary

Select only relevant tools per turn instead of sending all 24+ definitions. Two tiers: rule-based heuristics + Haiku for ambiguous cases.

Requirements

• Fast pass: keyword/intent heuristics map messages to tool categories
• LLM pass: Haiku evaluates message + tool names/descriptions, returns relevant subset
• Only selected tools' full schemas sent to brain
• Always include base fallback set (execute_shell, do_task)
• Depends on: usage data from meta-cognition (Implement meta-cognition (pattern recognition + habit formation) #9) for smart selection

References

• ARCHITECTURE.md — Tool selector (context shaping)