- config.py: E13 executor_thinking='off' to match the SEAM paper run
- run_ablate12.sh: dedicated E13 block (min_level=0 full pool, chunk=128,
  reward-trunc-penalty=0, eval R=1/T=0) reproducing the SEAM run config
- include the code-task/reflexion pipeline modules E13 imports at load time
  (main/trainer top-level import code_task/data_code/eval_reflexion)

…form)

- train_skill_v2.py: SKILL_GEN_FREEFORM/REGEN_FREEFORM_SYSTEM (a 'menu' prompt
  letting the skill model choose whatever form helps this problem — analysis,
  concept, pitfall, tiny example, blunt directive, even 'let's think step by
  step'), wired into style dispatch + --skill-style choices; the freeform prompt
  carries <skills></skills> wrapper examples so open-form outputs stay parseable
- config.py: E21 = bnpo/view-B/freeform (thinking on; see comment for why not off),
  STYLES + RUN_ORDER updated, self-check passes
- trainer.py: freeform shares narrative's 1100-char len budget

The BNPO family returned an already-normalized per-group token-mean with
num_tokens=0, so the framework's PER-TOKEN-MEAN path equal-weighted micro/dp
groups -> a double average (group token-mean, then equal weight over groups)
that sits between token-mean and sequence-mean and biases toward short
responses (degrades to pure sequence-mean as groups multiply). This diverged
from verl/SEAM's true token-mean and was non-orthogonal to skill-length study.

- grpo.py: BNPOLoss gains token_mean_scope='global'(default)|'micro'. 'global'
  returns the token SUM and reports num_tokens=sum(mask), routing into the
  framework SUM-loss path -> exact global token-mean, invariant to how the batch
  is split. 'micro' preserves the old behavior to reproduce E1-E20. Added a
  _loss_num_tokens hook (default 0) so GRPO/DRGRPO/OPSD are untouched. No public
  interface change; downstream grad + metric already branch on num_tokens.
- tests/loss/test_bnpo_token_mean.py: assert global is split-invariant (==true
  token-mean), micro reproduces the biased double-average, SEAM inherits global.
- run_ablate12.sh: E13(140G)/E21(80G) OOM'd in train forward at micro=8; set
  per-arm train_micro_batch defaults (E13=2xdp, E21=1xdp). The global token-mean
  fix makes shrinking micro mathematically equivalent, so effective batch and
  comparability are unchanged.

E13/E21 blocks referenced $TRAIN_MICRO_BATCH directly; under set -u an unset
env aborts with 'unbound variable'. Use ${TRAIN_MICRO_BATCH:-}.

e18_collect_kod.py 把 cookbook/human 加进 sys.path 后 import e23_rubric，
但该目录此前 0 文件入库，导致新机器 clone 后启动即
ModuleNotFoundError: No module named 'e23_rubric'。

依赖链已静态遍历确认闭合：
  e18_collect_kod -> e18_{kodcode,multidiag,prompts,select}
                  -> e23_rubric -> e23_prompts
e23_bcb 由 e18 其他脚本引用，一并提交。

judge 的失败是静默的：pytest 缺失不会让进程崩，只会让每题判 incorrect，
表现为 baseline_accuracy=0 / n_wrong=64/64 / collected 恒 0。B 机因此白跑
10 小时 65 个 chunk。

自检照抄 e18_kodcode.run_tests 的真实结构（solution.py + test_solution.py
+ _run.py + subprocess/sys.executable），只有走同一条路径，'通过'才等价于
judge 会判通过。同时检查依赖版本、教师 API key、分片参数、resume 种子。

…kers

- sanitize sys.argv around LLMAgent construction: ms-agent's
  Config.parse_args() mis-parses a foreign argv in forked/Ray workers
  (assert crash on value tokens; silent flag mispairing otherwise)
- tolerate ms-agent >= 1.6.0 API changes: prepare_skills ->
  _ensure_auto_skills(), dropped ms_agent.hooks and
  _append_task_notifications, defensive ToolResult field forwarding

… into feat/challenger