Both signals were tested against the public benchmark and neither held up:
the closed-form ETA is ill-posed at target_error=0 (the dominant real task
class) and fired on a tiny fraction of trajectories; gain margin is undefined
for the single-iteration majority, so it could only describe the struggling
minority. Rather than ship signals we don't trust, they're removed.

Breaking: removes LoopGainResult.gain_margin / first_eta_prediction /
first_eta_at_iteration, the LoopGain.gain_margin / LoopGain.eta properties,
and the gain_margin / first_eta_* telemetry fields (schema -> v4). The Aβ
trajectory (convergence_profile) and the five-band classifier are unchanged.
Reasoning preserved in loopgain-bench/LESSONS.md.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>