The Euler-Maruyama SDE in compute_inner_product used a complex-valued Wiener
increment dW = (N(0,1) + i·N(0,1))·√dt, which has E[|dW|²] = 2·dt — twice
the standard Wiener process. This caused coherence |T_τ|² to drift above 1.0,
making the collapse detector epistemologically invalid.
Two structural fixes:
1. Replace complex dW with real dW: E[dW²] = dt (correct Wiener energy)
2. Renormalize similarity to unit circle after each GBM step, enforcing
|T_τ|² ≤ 1 as a hard invariant rather than relying on downstream clipping
Also derive dt from token_freq (default 0.05s at 20Hz) instead of the
hardcoded dt=1.0 that ignored all hardware clock configuration.
Adds tests/test_falsification.py: 12-test falsification harness proving the
defect via Monte Carlo (100k samples, E[|dW_complex|²]/dt ≈ 2.0) and
verifying the patch produces 0 violations across 10,000 engine steps.
Adds data/telemetry_sample.json: 300 synthetic records (GPU + Pi Zero)
confirming coherence>1 violations appear in both hardware environments.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>