[reward] fix: keep boxed answers within the documented length budget

[JimmyWang0417]

What does this PR do?

compute_score() already limits math-DAPO solutions to their final 300
characters, matching the documented maximum MATH-500 answer length of 159
characters. is_correct_strict_box() then applies another 100-character slice,
which can remove the opening \boxed{ marker from a valid answer and silently
score it as incorrect.

This change keeps the existing 300-character bound as the single window. Strict
boxed-answer extraction receives the full bounded string while preserving the
existing four-element validation for pause_tokens_index.

Checklist Before Starting

• Searched open and closed PR/issue text for math_dapo,
  is_correct_strict_box, boxed-answer length, and pause_tokens_index work.
• Performed a complete paginated file-level scan of all open PRs.
  • Open PR [training_utils] fix: improve null handling and type safety in math_dapo #4015 changes null handling, return types, and branching in the same
    function but leaves both 100-character slices unchanged.
  • Open PR [trainer] feat: add rollout saving and more training log to DAPO #2323 only adds ground_truth to the returned score dictionary.
  • Open PR Routing replay implemented with SGLang and FSDP #4443 broadly replaces the scorer for unrelated routing-replay work;
    it does not claim or test this length-window defect.
  • Open PR [Fix][DAPO] Fix strict_box_verify not being forwarded to DAPO compute_score. #4957 propagates strict_box_verify through reward managers without
    changing math_dapo.py.
• Validated the title with tests/special_sanity/check_pr_title.py.

Test

Before the fix, the documented-length strict, fallback, and pause-index cases
failed because the second 100-character slice removed the box opener:

PYTHONDONTWRITEBYTECODE=1 python /tmp/verl_w9_test_runner.py \
  '/home/jimmywang0417/Git Repositories/cv/.two-agent-cooperation/verl-w9'
3 failed, 2 passed

After removing only the redundant slice assignments:

PYTHONDONTWRITEBYTECODE=1 python /tmp/verl_w9_test_runner.py \
  '/home/jimmywang0417/Git Repositories/cv/.two-agent-cooperation/verl-w9'
5 passed

PYTHONDONTWRITEBYTECODE=1 python -m compileall -q \
  verl/utils/reward_score/math_dapo.py \
  tests/utils/reward_score/test_math_dapo_on_cpu.py
Passed

uvx --with hydra-core pre-commit run --all-files \
  --show-diff-on-failure --color=never
All 13 configured hooks passed

The full dependency test environment and remote CI were not run locally.

API and Usage Example

No signature, return-schema, configuration, or score-value changes. Existing
callers continue to use compute_score() unchanged.

Design & Code Changes

• Keep compute_score() as the sole owner of the 300-character efficiency
  window.
• Remove the two inner 100-character slice assignments before strict box
  extraction.
• Preserve the four-element pause_tokens_index validation without inferring an
  unproven token-to-character coordinate conversion.
• Add focused CPU coverage for strict mode, fallback mode, final-box selection,
  pause-index compatibility, and unchanged no-box failure behavior.

AI Assistance

OpenAI Codex was used for repository exploration, implementation assistance,
test orchestration, and review. I reviewed every changed line, can explain the
change end-to-end, and personally ran the focused regression tests.

Checklist Before Submitting

• Read the Contribute Guide and AGENTS.md.
• Applied the full-repository pre-commit checks.
• Added a CPU regression test collected by cpu_unit_tests.yml.
• Documentation is not required because no public interface or configuration
  changes.
• Request project CI in Slack or Feishu after the PR is available.
• The recipe submodule is not involved.