Add dummy loss guard and policy output key registration

[relh]

Summary

• add a core dummy-loss guard for unused policy outputs and remove per-loss shims
• add Loss.policy_output_keys and union used keys in the train loop
• refine ActionSupervised action spec handling for discrete vs continuous actions

Testing

• not run (not requested)

[relh]

• Revert hyperparam defaults and eval cadence #4485
• Fix supervisor teacher behavior and legacy BC mode #4484
• Rename HyperUpdateRule to ScheduleRule #4483
• Add dummy loss guard and policy output key registration #4482 👈 (View in Graphite)
• main

This stack of pull requests is managed by Graphite. Learn more about stacking.

[datadog-official]

✅ Tests

🎉 All green!

❄️ No new flaky tests detected
🧪 All tests passed

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 3044eea | Docs | Was this helpful? Give us feedback!}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".