[faresobeid]

Two small cleanups in orchestrator/algo/. Item 4 of the orchestrator review.

• action_loss_type is now declared once, on the algorithm config classes. Before, it was declared on the config classes and again on the runtime classes (algo/base.py, sft.py, opd.py, opsd.py), and build_algorithm asserted that the two agreed. Now Algorithm.__init__ copies it from the config, so train_sink's env.algorithm.action_loss_type is unchanged.
• Algorithm.finalize_group is removed. It only called score_group, and train_sink now calls score_group directly. finalize_episode is kept: its "has sampled tokens" guard matters for echo, whose score_episode does not filter on the mask.
• docs/algorithms.md is updated to match.

Behavior change: none. Every algorithm type gets the same loss routing as before, because the old assert guaranteed that the config and runtime values were equal.

Checks

• pytest tests/unit/orchestrator tests/unit/test_configs* -m "not gpu": 291 passed (test_algorithms covers the per-type action_loss_type).
• ruff check, ruff format --check: clean.

GPU checks: none needed.

Conflicts

algo/base.py and algo/__init__.py: #3802, #3803 (small). #3838 (fold GenerationSource) also edits algo/base.py and algo/__init__.py (connect_frozen_client), which needs a trivial rebase. train_sink.py:245: #3652 rewrites train_sink.py (a one-line change there).

---

Note

Low Risk
Refactor-only: loss routing and group scoring call paths are unchanged in effect; no auth, data, or trainer hot-path logic changes.

Overview
Single source of truth for loss routing: runtime Algorithm instances now take action_loss_type from the algorithm config in __init__ instead of duplicating it on each subclass and asserting agreement in build_algorithm. Per-algorithm class variables on OPD, OPSD, and SFT are removed; train_sink still reads env.algorithm.action_loss_type when stamping samples.

Thinner scoring API: Algorithm.finalize_group is dropped and the train sink calls score_group directly when a group has trainable survivors. finalize_episode remains as the guarded entry point for rollout-local score_episode (skips episodes with no trainable sampled tokens).

Docs: algorithms.md is updated to describe this pipeline wiring and config-owned action_loss_type.

No intended training behavior change—loss routing values are the same as before when config and runtime matched.

^{Reviewed by Cursor Bugbot for commit 1ebee7f. Bugbot is set up for automated code reviews on this repo. Configure here.}