feat: wire Query→NARS feedback loop with BF16 causal learning

Closes the missing feedback circuit between recognition results and NARS
truth value updates. The BF16 1+7+8 prefix decomposition (sign/exp/man)
maps directly to NARS inference rules:
- sign flips → NarsInference::negation() (class-level cause)
- exponent shifts → NarsInference::abduction() (magnitude cause)
- mantissa noise → filtered out (irrelevant for learning)

The feedback module wires:
- EnergyConflict → TruthValue (agreement_ratio → frequency)
- LearningSignal → evidence history ring buffer (W160-W175)
- Hybrid attention weights → WideMetaView (W144-W159)
- RL Q-values update from awareness reward (W32-W39)

19 tests, all passing.

https://claude.ai/code/session_016HQG2eoyNAwML6WP8WhbAG

Author: claude