Phase 12 finding 4: step_emb row 4 is a zero vector (never trained; the clamp
reuse made the 16-step depth collapse mechanical). jitter p=0.15 per batch
element per step samples a uniform random row; eval unchanged (deterministic
clamp). Verified: 5/5 rows receive grad at B=32, baseline row4 grad=0, gates 12/12.
bench_learn: set logp_floor/kl_hard/min_updates attrs expected by train_step after logp_floor fix