perf: 5.7x Doppler extraction speedup, trust kill switch, fix NN benchmark

Optimization:
- Cache mean phase per frame in ring buffer for O(1) Doppler access
- Sliding window (last 64 frames) instead of full history traversal
- Doppler FFT: 253.9us -> 44.9us per frame (5.7x faster)
- Full pipeline: 719.2us -> 254.2us per frame (2.8x faster)

Trust kill switch:
- ./verify: one-command proof replay with SHA-256 hash verification
- Enhanced verify.py with source provenance, feature inspection, --audit
- Makefile with verify/verify-verbose/verify-audit targets
- New hash: 0b82bd45e836e5a99db0494cda7795832dda0bb0a88dac65a2bab0e949950ee0

Benchmark fix:
- NN inference_bench.rs uses MockBackend instead of calling forward()
  which now correctly errors when no weights are loaded

https://claude.ai/code/session_01Ki7pvEZtJDvqJkmyn6B714

v1/data/proof/expected_features.sha256

@@ -1 +1 @@
-7b9ed15a01a2ae49cb32c5a1bb7e41361e0c83d9216f092efe3a3e279c7731ba
+0b82bd45e836e5a99db0494cda7795832dda0bb0a88dac65a2bab0e949950ee0