README: crossing costs once the observation's bytes (E46), other trainers (E48, E49), the boundary is transparent (E50, E52, E53); E49's figure

[tactino]

The README's table of answers had stood still since 09-29. Two of its rows were wrong by now, and three results were missing.

Corrected

• What does crossing cost? The answer now comes from E46: 11.6 ms for 184 KiB at 23.7 MB/s, with each observation crossing once under reuse-feedback-obs, which the server offers and the env client uses by default. The old answer was E43's "twice the observation's bytes, 21 ms at 18 MB/s". E43 is still linked as the measurement before the feature.
• Is that cheap beside a VLA forward pass? On that link it is 12% of pi0.5's 100 ms forward, not 21%.

Added

• E49 (with E48): RLinf, Stable-Baselines3 and CleanRL train PlugRL env clients through plugrl-bridges, 3 of 3 seeds in every pairing.
• E50 (with E52 and E53): SB3 ends on byte-identical weights with its environments behind the protocol.

Also

• The count of experiments no longer names a number, which went stale with every new experiment.
• experiments/README.md linked https://github.com/PlugRL/plugrl-bridges, which is private and 404s for everyone else. It now names the repository as not public yet. Its E10 row adds E46's 12%.
• experiments/e49-multi-trainer/figure.py draws every seed's curve from the bridge's episode logs (results/registered/*/episodes.csv). It makes the figure on the project page (plugrl.github.io#25). The script passes ruff check and ruff format, and two runs give byte-identical PNGs.