[FEATURE] Repetition filter operator for degenerate rollout responses

[KunyangZhang]

Motivation

#470 invites new ExperienceOperator implementations, including "an operator to filter out
low-quality experiences based on predefined criteria". The existing filters cover reward thresholds
(reward_filter), zero-variance groups (reward_std_filter), DAPO dynamic sampling
(dapo_dynamic_sampling), truncated responses (mask_response_truncated) and invalid rewards
(invalid_reward_filter).

None of them looks at the text of the response. Repetition collapse is a common failure mode of
on-policy rollouts — especially for smaller models on long-horizon agent tasks: the policy keeps
emitting the same pattern instead of terminating. Those trajectories consume rollout budget and give
almost no learning signal, yet they currently enter the buffer like any other experience.

Proposal

A repetition_filter operator that drops experiences whose response has a high duplicate n-gram ratio.

explorer:
  experience_pipeline:
    operators:
      - name: repetition_filter
        args:
          ngram_size: 4          # sliding window used to measure repetition
          max_repeat_ratio: 0.5  # drop above this duplicate ratio
          min_response_tokens: 32  # shorter responses are always kept (ratio too noisy)

• Only the model-generated part of the response is inspected: tokens before prompt_length are
  ignored, and when action_mask is present (multi-turn experiences) observation tokens are ignored
  as well.
• The response is dropped when duplicate_ngrams / total_ngrams > max_repeat_ratio; for example
  [7, 8, 9] * 20 scores 0.94 with ngram_size=4, while a diverse response scores 0.0.
• Metrics: filtered_count, skipped_short_count, skipped_empty_count, repeat_ratio/mean,
  repeat_ratio/max.
• Implemented on ExperienceOperatorV1 (the async interface the operator guide asks new operators to
  use), registered as repetition_filter in EXPERIENCE_OPERATORS, with unit tests covering the
  ratio function, the filter, the action_mask path and registry creation.

Question before I open the PR

1. Is a repetition/degeneracy filter in scope for the buffer, or would you rather see this handled
   inside Workflow/Algorithm?
2. Dropping the experience is the simplest behaviour and matches the wording in Good First Issue List #470. An alternative
   is to keep the experience but zero its action_mask (like mask_response_truncated does), so the
   rollout still counts for reward statistics. Which do you prefer?
3. Defaults: I picked ngram_size=4, max_repeat_ratio=0.5, min_response_tokens=32. Happy to
   change them or drop the length guard.

I have the implementation and tests ready locally (CPU-only, pytest tests/buffer/repetition_filter_test.py
→ 11 passed) and will open the PR as soon as you confirm the direction.