feat: add DAPO asymmetric GRPO clipping

[0z5a]

Summary

• add independent lower and upper GRPO clipping epsilons for DAPO Clip-Higher
• preserve the existing symmetric clip_eps=0.2 behavior when the new options are unset
• expose --grpo_clip_eps_low and --grpo_clip_eps_high through CLI/YAML configuration
• reject non-finite, negative, inverted, or invalid lower-bound settings
• document the asymmetric objective and the DAPO 0.20/0.28 example

Objective

The importance-ratio clamp becomes:

clip(rho, 1 - clip_eps_low, 1 + clip_eps_high)

Setting both values to 0.2 is exactly the previous GRPO objective. Setting clip_eps_low=0.2 and clip_eps_high=0.28 selects the Clip-Higher component described in the DAPO technical report, without implicitly enabling DAPO's separate dynamic-sampling, token-aggregation, or overlong-reward components.

L20 microbenchmark

Exact implementation on NVIDIA L20 (SM89), PyTorch 2.11.0+cu128, CUDA 12.8. One loss vector contains 1,048,576 FP32 token ratios/advantages; 30 warmups, then A-B-B-A with 20 samples per path and 100 iterations per sample.

clamp	median	p90
symmetric 0.20/0.20	0.05333 ms	0.05406 ms
DAPO 0.20/0.28	0.05288 ms	0.05428 ms

The −0.8% median difference is measurement noise: both paths launch the same operations and differ only in clamp constants. The feature adds no new tensor pass to the GRPO loss hot path.

Validation

• numerical formula test covers positive upper clipping and unchanged negative lower clipping
• symmetric-default compatibility test
• finite/range/order validation tests
• focused Ruff + pytest: 14 passed
• CUDA_VISIBLE_DEVICES=5 bash scripts/pre_commit.sh --skip-deps: 655 passed

[0z5a]

Superseded by #58, which preserves the same Clip-Higher configuration and compatibility tests while adding token/sequence aggregation, overlong shaping, and the complete L20/InfraSWE evidence. Keeping one review target avoids reviewing duplicate objective code. The commit and discussion here remain in history.