Record RTX 4090 action-norm performance profile

[Al-does]

Summary

• record a controlled Vast RTX 4090 baseline for the action-norm experiment
• document the measured bottlenecks and every tested configuration ablation
• record the matched end-to-end delta from the promoted RL-Harness hot-path optimizations
• retain quality evidence for rejecting the faster three-epoch PPO schedule

Results

• environment microbenchmark: 8.167 s → 5.510 s per 100k steps (-32.5%)
• matched PPO throughput: 12,537 → 13,049 env steps/s (+4.08%)
• matched PPO iteration wall time: 5.255 s → 5.048 s (-3.93%)
• matched seeded reward/KL/loss/value metrics remained identical

Dependency

The promoted implementation is on Al-does/RL-Harness branch cursor/optimize-transition-validation-1446 (commits 740c761 and 1ba1ee3).