feat: add configurable DAPO objective controls

[0z5a]

Summary

• add DAPO Clip-Higher through independent lower and upper GRPO clipping bounds
• add configurable token-level versus sequence-level policy-loss aggregation
• add soft overlong reward shaping with buffer and scale controls
• preserve the existing objective exactly when all new options remain at defaults
• include a reproducible L20 benchmark and frozen InfraSWE comparison evidence

Dynamic sampling is intentionally not included in this PR: correct refill semantics require a rollout-buffer lifecycle change rather than silently discarding zero-variance groups.

L20 benchmark

Environment: NVIDIA L20 (SM89), PyTorch 2.11.0+cu128, CUDA 12.8, seed 3407, 20 warmups and 100 timed repetitions.

Shape (batch × group × response)	Baseline median	Default candidate	Delta	Full DAPO median	Absolute DAPO cost
4 × 8 × 256	0.256 ms	0.254 ms	-0.80%	0.312 ms	+0.056 ms
4 × 8 × 1024	0.257 ms	0.257 ms	0.00%	0.316 ms	+0.059 ms
8 × 8 × 2048	0.269 ms	0.269 ms	0.00%	0.330 ms	+0.060 ms
4 × 8 × 4096	0.269 ms	0.268 ms	-0.38%	0.329 ms	+0.059 ms

Default-path maximum absolute output difference is 0 for all four cases. Timings isolate the objective calculation and exclude model forward/backward work.

Validation

• local suite: 635 passed, 103 skipped
• L20 focused suite: 26 passed
• Ruff format/import checks: passed
• InfraSWE diagnostic ProjectFit: 92.15
• InfraSWE BenchmarkTrust: 97.40
• official InfraSWE score: unresolved because signed draft-seal, hidden probes, fresh-process replay, and verified evidence-manifest inputs are unavailable

The implementation follows the objective components described in the DAPO paper. Raw benchmark JSON, the runnable benchmark, and InfraSWE scoring inputs are committed under benchmarks/.