[codex] align hosted training config examples

[tim0120]

Summary

• update SFT and OPD cookbook configs to match the hosted CLI distillation surface
• add loss = "opd" and remove unsupported public teacher knobs like teacher_tau, save, and replay
• route training environment overrides through env.args so examples validate with the hosted train config parser
• align the related RL continuation examples/docs with the same Hosted Training env schema

Why

The new OPD setup uses the public hosted CLI shape (loss = "opd" plus [teacher]) while PrimeRL/runtime handles teacher endpoint wiring internally. The cookbook should document that hosted surface instead of PrimeRL internal config fields or stale env override tables.

Hosted Training uses one env config schema across rl, sft, and opd: per-env taskset/harness overrides are passed through env.args, not top-level [env.taskset] or [env.harness] tables. Keeping the related RL examples on the old shape would leave cookbook training examples that the new parser rejects.

Validation

• python3 + tomllib: parsed all 38 cookbook TOML files
• uv run python: loaded changed hosted training TOMLs with prime_cli.commands.rl.load_config from the local OPD CLI branch
• git diff --check

Related

• PrimeRL: fix: harden OPD teacher logprob scoring prime-rl#2683
• Prime CLI: APR-57: add hosted OPD CLI config prime#699