Repo cleanup: remove run logs and forwarding stubs, single source for pins, merge weight and lag bounds

docs/MILES_RL.md

@@ -775,7 +775,7 @@ identity keeps the normal in-phase recovery behavior instead.
 ## Engine selection
 
 `--rl-engine {legacy,ports}` selects how an island drives Miles. It is accepted
-by `yeto launch`, `python3 -m yeto.rl.learner`, the SSH harness (through the
+by `yeto launch`, `python3 -m yeto.rl.adapters.miles.island_entry`, the SSH harness (through the
 launch arguments), `scripts/benchmark_rl.py` and `yeto-rl-export`. The default
 is `ports`; the choice is never inferred from the model, island count or other
 flags, is fixed before startup, and does not change during a run.
@@ -851,15 +851,15 @@ it joins outer sync, and refuses to start if the hash differs.
 Where the spec comes from:
 
 1. `--rl-algorithm-spec PATH`, a v1 or v2 JSON file. `yeto launch` and
-   `python3 -m yeto.rl.learner` both accept it; it applies to `ports` only.
+   `python3 -m yeto.rl.adapters.miles.island_entry` both accept it; it applies to `ports` only.
 2. Without that flag, the legacy CLI builds the spec exactly as in R0.
 3. Mapped Miles flags found in the extra argv are absorbed into the spec.
 
 A spec that only uses R0 fields keeps the R0 v1 canonical JSON, so its hash
 is unchanged. Setting any new field switches the spec to
 `yeto-rl-algorithm-spec-v2`.
 
-**Absorb and reject.** `yeto/rl/engine/miles_adapter/algorithm_flags.py` maps
+**Absorb and reject.** `yeto/rl/adapters/miles/algorithm_flags.py` maps
 each Miles flag to a spec field. Every flag in that table is adapter-owned:
 
 - A mapped flag in the extra argv is absorbed. It is recorded as
@@ -986,7 +986,7 @@ algorithm change declares it after its single-GPU smoke passes.
 `--rl-allow-unverified-mechanism DIMENSION:NAME` (repeatable, for example
 `features:clip_higher`) exempts only the named mechanisms from the "not
 declared" check, and only on a single-island run **without outer sync**.
-Every launched run and every `yeto.rl.learner` run joins a syncer, so today
+Every launched run and every `yeto.rl.adapters.miles.island_entry` run joins a syncer, so today
 the allowance is refused on those entry points; only the fake engine and the
 dry run (which models a sync-less single island) accept it.
 Every other check still applies. The allowance does not change the hash. It is
@@ -1714,7 +1714,7 @@ $M --dry-run --extra "--advantage-estimator ppo --gae-variant cross_segment_per_
 
 ### Port responsibilities
 
-| port | responsibility | Miles adapter (`yeto/rl/engine/miles_adapter/`) |
+| port | responsibility | Miles adapter (`yeto/rl/adapters/miles/`) |
 | --- | --- | --- |
 | `RolloutPool` | generate one rollout of complete groups from the published policy; members | `rollout.py` over `InferenceController` + `RolloutExecutor`; metadata from `rollout_meta_hook.py` inside the rollout process |
 | `TrainerGroup` | one optimizer step on an opaque batch; onload/offload; grad norm for the per-round invariant | `trainer.py` over the single-cell actor group |
@@ -1933,7 +1933,7 @@ it is rejected as undeclared without the allowances and accepted with them.
 
 ## Algorithm status matrix
 
-"Declared" means the Miles adapter (`yeto/rl/engine/miles_adapter/entry.py`,
+"Declared" means the Miles adapter (`yeto/rl/adapters/miles/entry.py`,
 `MILES_DECLARED`) accepts the mechanism without
 `--rl-allow-unverified-mechanism`, which requires GPU evidence that it takes
 effect. Declarations say nothing about training gains. G1 = single island on
@@ -2161,8 +2161,8 @@ The implementation is intentionally confined to the RL boundary:
 | --- | --- |
 | `yeto/rl/core.py` | canonical LoRA, deterministic RL fragments, policy snapshots |
 | `yeto/rl/decoupled.py` | raw anchors, BCAST/PULL/PUSH state, final cut and benchmark consolidation |
-| `yeto/rl/miles.py` | rollout-token validation, safe-boundary hook, island checkpoint |
-| `yeto/rl/learner.py` | Miles argument mapping and runtime configuration |
+| `yeto/rl/adapters/miles/legacy/engine.py` | rollout-token validation, safe-boundary hook, island checkpoint |
+| `yeto/rl/adapters/miles/island_entry.py` | Miles argument mapping and runtime configuration |
 | `yeto/rl/export.py` | authoritative checkpoint to standard PEFT |
 | `yeto/rl/initial_adapter.py` | validated Decoupled PEFT policy warm start |
 | `scripts/benchmark_rl.py` | equal-work native/strict/decoupled comparison |

docs/WANDB.md

@@ -81,7 +81,7 @@ count — so a curve can be filtered by cloud, region, or GPU.
 
 ## RL islands
 
-An RL island runs MILES behind `yeto.rl.learner`, so there is no inner
+An RL island runs MILES behind `yeto.rl.adapters.miles.island_entry`, so there is no inner
 loop here to instrument — but the island already keeps its own structured
 event tape (`--event-tape`, written by `_append_rl_event`), carrying every
 rollout round, fragment broadcast, fragment push, and strict-invariant

docs/research/rl-algorithms/research.md

@@ -50,7 +50,7 @@
    - 但 `grpo` 和 `gspo` 的 advantage 只把 reward 广播到每个 token，KL 被丢弃（`LH/advantages.py:53-57`；`LH/math_utils.py:462-469` 的 `get_grpo_returns`）。
    - KL 只有在 `--use-kl-loss` 时才进入 loss（`LH/losses.py` 的 `use_kl_loss` 分支）。
    - 因此 `kl_coef>0` 加 GRPO 的效果是多占一份 ref 模型的显存和前向算力，梯度没有任何变化。
-   - legacy 的 `expert_full` 分支固定传 `--kl-coef 0.001`（`yeto/rl/learner.py:1068`）。legacy 用的是 agentenv fork，那边的 GRPO 是否同样忽略 KL **[未核实]**。
+   - legacy 的 `expert_full` 分支固定传 `--kl-coef 0.001`（`yeto/rl/adapters/miles/island_entry.py:1068`）。legacy 用的是 agentenv fork，那边的 GRPO 是否同样忽略 KL **[未核实]**。
 
 5. **在 ports 的串行 driver 下，外层同步不会让 rollout 数据变旧** [源码]：
    - 每轮的顺序是：用已发布策略生成，训练一次，在安全点 apply cut，再发布（`driver.py:1-10, 418-443`）。

evidence-forkg1/s13-forkg1-modal.sh

@@ -66,8 +66,8 @@ cfg.validate()
 keys = ("app_name", "function_name", "gpu_request", "num_nodes", "gpu_exact", "retries", "timeout_s",
         "tape_volume_name", "tape_subdir", "cpu_request", "memory_request_mib", "image_ref")
 print("modal cfg:", {k: getattr(cfg, k, "<absent in this code>") for k in keys})
-learner = next(l for l in task.run.splitlines() if "-m yeto.rl.learner" in l)
-print("learner flags:", learner.split("-m yeto.rl.learner", 1)[1].strip())
+learner = next(l for l in task.run.splitlines() if "-m yeto.rl.adapters.miles.island_entry" in l)
+print("learner flags:", learner.split("-m yeto.rl.adapters.miles.island_entry", 1)[1].strip())
 print("algorithm sha:", args.rl_expected_algorithm_sha256)
 print("overlay:", getattr(args, "rl_miles_overlay", None)); import re as _re
 print("setup overlay lines:", [l.strip()[:160] for l in (task.setup or "").splitlines() if _re.search(r"REFUSING|sha256sum|git apply|overlay", l)][:12])

examples/custom_reward/README.md

@@ -4,7 +4,7 @@
 
 - 按名称：`import examples.custom_reward.reward` 后名称 `ends_with_label` 已注册。
 - 按引用：`examples.custom_reward.reward:ends_with_label`。
-- Miles 下取 `--custom-rm-path` 的值：`yeto.rl.engine.miles_adapter.rewards.miles_custom_rm_path(<名称或引用>)`；名称未注册时报错并列出已注册名称。
+- Miles 下取 `--custom-rm-path` 的值：`yeto.rl.adapters.miles.rewards.miles_custom_rm_path(<名称或引用>)`；名称未注册时报错并列出已注册名称。
 - 加载时记录插件身份（模块、函数名、源码文件 sha256，见 `yeto.rl.rewards.registry.load_reward`），改源码身份即变。
 
 测试：`python -m pytest examples/custom_reward/test_reward.py`。

openspec/changes/add-nebius-verda-modal-clouds/run-logs/EVIDENCE.md

@@ -0,0 +1,24 @@
+# Evidence moved out of the repository
+
+Run logs that used to live under this directory were removed from git by change `rl-code-quality-cleanup` (2026-10-12).
+Copies are kept outside the repository at:
+
+    /home/michael/work/infra-drafts/evidence/add-nebius-verda-modal-clouds/run-logs/
+
+The relative path under that directory is the same as the old path under this directory.
+Moved files:
+
+- task-8.7-cleanup-verification.txt
+- task-8.7-command.sh
+- task-8.7-modal-island-metrics.txt
+- task-8.7-syncer-tape.txt
+- task-9.2-budget60-post-change.json
+- task-9.2-budget60-post-change.stderr
+- task-9.2-budget60-pre-change.json
+- task-9.2-budget60-pre-change.stderr
+- task-9.2-command.txt
+- task-9.2-post-change.json
+- task-9.2-post-change.stderr
+- task-9.2-pre-change.json
+- task-9.2-pre-change.stderr
+- verify_cleanup.sh