feat: hardcode rollout filters and zero-advantage drop

[mikasenghaas]

Summary

Orchestrator cleanup: batching stays batch_size + group_size, and everything between scoring and training is hardcoded.

• Hardcode gibberish/repetition monitoring. The composable filter config (pre_batch_filters / post_batch_filters, discriminated FilterConfig union, per-slot enforce) is gone. filters.py is now three hardcoded checks with the old default thresholds as module constants. Gibberish and repetition detection are monitor-only: tracked per rollout in metrics, never dropped.
• Hardcode zero-advantage filtering as a pre-batch drop. A rollout whose advantage stream is all zero is dropped before it takes a batch slot, so every batch fills with batch_size informative samples: the trained-on batch is predictable, and the per-step sampling time varies with the zero-advantage rate. orchestrator.count_zero_advantage_in_batch = true opts back into the old accounting (dropped rollouts still count toward batch_size — fixed sampling effort per step, variable trained-on samples). Echo keeps its zero-advantage rollouts via a new Algorithm.trains_on_zero_advantage class flag (its ce component trains observation tokens regardless of credit); opd/opsd assign no advantage and never match. The sink warns every 25 consecutive all-zero-advantage groups so a stalled batch fill is visible.
• Simplify batching config. token_batch_size (the alternative rollout-count-free mode) and oversampling_factor (a max_inflight_episodes derivation) are removed. max_inflight_episodes defaults to batch_size; set it explicitly to oversample — configs that used a factor now carry the resolved value.
• Flatten the check-verdict metric keys. {scope}/{subset}/{agent}/filters/<name>/mean → {scope}/{subset}/{agent}/<name>/mean, next to the other per-trace verdicts (is_trainable, is_filtered).
• Docs and the monitor-run skill are updated to match; the README also drops its AIPO mention.

Breaking

• orchestrator.pre_batch_filters / orchestrator.post_batch_filters are removed. Gibberish/repetition run always in monitor mode; enforcement is no longer available (examples/advanced/glm-5.2/swe.toml previously enforced gibberish and now only monitors it). Zero-advantage dropping is always on; the old default budget accounting is available via count_zero_advantage_in_batch = true. Echo no longer needs a filter override.
• The zero-advantage default moves from post-batch to pre-batch: dropped rollouts no longer consume batch slots, so a step samples more rollouts but always trains on batch_size informative samples.
• orchestrator.token_batch_size is removed; batch_size is the only batching unit.
• orchestrator.oversampling_factor is removed. Set max_inflight_episodes = batch_size x factor explicitly instead.
• W&B keys pre_filters/all/dropped_rate and pre_filters/all/<name>/rate are gone, and the check verdicts drop their filters/ segment (.../filters/gibberish/mean → .../gibberish/mean).

Verification

• tests/unit (configs incl. parsing every checked-in TOML, orchestrator, utils) passes; test_qwen3_vl_e2e fails identically on main (pre-existing).
• 5-step smoke run of configs/ci/integration/reverse-text/start.toml on 2 GPUs: full 128-rollout batches every step, reward improves 0.13 → 0.25, no errors, clean exit.

🤖 Generated with Claude Code