chore: finish the legacy-ReFL deletion sweep

.github/CODEOWNERS

@@ -40,12 +40,10 @@
 /unirl/train/                           @CjhHa1 @celve @haonan3
 /unirl/train/backend/                   @celve @zzhuoxin1508 @CjhHa1
 /unirl/train/stack/                     @leviking98z-rgb @haonan3
-/unirl/train/refl/                      @celve @Ideny42
 /unirl/train/sft/                       @haonan3 @xshrz
 /unirl/train/unified_model_stack.py     @Ideny42 @zzhuoxin1508
 /unirl/trainer/                         @celve @Ideny42 @CjhHa1
 /unirl/train_sft.py                     @haonan3 @xshrz
-/unirl/train_refl.py                    @celve @Ideny42
 /unirl/train_pe.py                      @Jayce-Ping @Ideny42
 /unirl/train_unified_model.py           @Ideny42 @zzhuoxin1508
 

experimental/refl/README.md

@@ -50,6 +50,19 @@ environment, not the code. Reward and actor share one Python process, so
 recipe `requirements.txt` files may only ADD packages, never re-pin the
 core stack.
 
+## Deliberately not ported from the legacy core path
+
+Cut on the tier's minimalism rule (features return only with a recipe that
+needs them; see the placement comment in `trainer.py`):
+
+- **Periodic eval** (`eval_interval` / `eval_sample` / `eval_cfg_text_scale` /
+  multi-reward `eval_rewards` suites, from #202) — the training reward curve is
+  this package's success signal; bring eval back via `unirl.trainer.eval_suites`
+  when a recipe needs checkpoint selection.
+- **`reward_fraction` cross-slab reward placement** — colocated-only here.
+- `param_checksum` debug probe, `activation_checkpoint_vae` knob (checkpointing
+  is unconditionally on in the grad decode paths).
+
 ## Verification
 
 | Config | Hardware | Head | Status |

experimental/refl/examples/wan21_t2v_videoalign_refl.yaml

@@ -18,7 +18,7 @@ max_grad_norm: 1.0
 actor:
   _target_: experimental.refl.roles.ReflActorRole
   # Family selector — swap pipeline_target + model_config for another family,
-  # no code changes (mirrors ReFLPolicy's pipeline_target contract).
+  # no code changes (the pipeline_target contract).
   pipeline_target: experimental.refl.models.wan21.Wan21ReflPipeline
   block_class_names: ["WanTransformerBlock"]
   # REFL loss: -(reward - baseline) / scale * weight + kl_weight * KL.

experimental/refl/roles.py

@@ -1,6 +1,6 @@
 """ReflActorRole — family-agnostic REFL/BPTT actor Remote for the refl recipe.
 
-Mirrors :class:`unirl.train.refl.policy.ReFLPolicy` (the SD3 image-ReFL actor):
+The family-agnostic REFL actor:
 a **config-chosen** ``Pipeline`` (``pipeline_target`` + ``model_config``, no
 per-family imports), FSDP-wrapped in place via ``FSDPBackend``, driven by three
 driver RPCs per step under the distributed ``enable_grad()`` context::
@@ -100,7 +100,7 @@ def __init__(
     def initialize(self) -> None:
         torch.cuda.set_device(self.device)
         # Default PG over the actor role's workers (env:// from Remote.setup's
-        # dist_env); FSDP2 fully_shard wraps over it. Same order as ReFLPolicy.
+        # dist_env); FSDP2 fully_shard wraps over it. Phase-0-validated order.
         if self.rank_info is not None and int(self.rank_info.world_size) > 1 and not dist.is_initialized():
             dist.init_process_group(backend="nccl")
 

experimental/refl/trainer.py

@@ -1,7 +1,6 @@
 """REFLTrainer — recipe driver for WAN REFL/BPTT (video reward backprop).
 
-The video sibling of :class:`unirl.trainer.refl.RewardBackpropTrainer`: two
-roles, always — a :class:`experimental.refl.roles.ReflActorRole` (FSDP WAN +
+Two roles, always — a :class:`experimental.refl.roles.ReflActorRole` (FSDP WAN +
 grad BPTT sampling + optimizer) and a frozen differentiable video reward
 (:class:`unirl.reward.service.RewardService`), colocated on the same worker
 slab so decoded video never leaves the GPU. Each step runs, under the

unirl/models/sd3/pipeline.py

@@ -136,7 +136,8 @@ def build_conditions(
     ) -> SD3Conditions:
         """Encode prompts (+ optional CFG negatives) into ``SD3Conditions``.
 
-        Shared by :meth:`generate` and the ReFL draft path (``draft_generate``).
+        Shared by :meth:`generate` and grad-sampling adaptations (e.g. the
+        experimental ReFL pipeline).
         Applies SD3's empty-negative default (diffusers parity) when CFG is on and
         no negative was supplied — see the rationale quoted in :meth:`generate`.
         """

unirl/trainer/eval_suites.py

@@ -25,7 +25,7 @@
 Placement: :func:`build_eval_suites` must be called inside the SAME placement
 context that created the trainer's training reward — each suite reward becomes
 a sibling remote there, so where the trainer has a ``reward_fraction`` slab
-(DiffusionTrainer, ReFL) ALL eval rewards share that dedicated-GPU slab, and
+(DiffusionTrainer) ALL eval rewards share that dedicated-GPU slab, and
 elsewhere (PE, UnifiedModel) they colocate with the training reward.
 
 Data: an own-set suite instantiates its own driver-side data source — the
@@ -34,10 +34,9 @@
 format the trainer's data source reads (txt / JSONL / JSON manifests with
 metadata) works per suite.
 
-Scoring uses the trainer's own reward interface: composed/rollout trainers call
-``suite.reward.score_and_attach``, ReFL calls ``score_differentiable`` — a
-suite's backend must support whichever its trainer uses (the same contract as
-the training reward).
+Scoring uses the trainer's own reward interface (every current consumer calls
+``suite.reward.score_and_attach``) — a suite's backend must support whichever
+its trainer uses (the same contract as the training reward).
 """
 
 from __future__ import annotations