fix: validate DAPO integer labels without float coercion

[morluto]

Problem

The DAPO scorer canonicalizes ground-truth labels with int(float(gt)). That conversion can change the reward result:

• "0.5" becomes "0", so a prediction of 0 is incorrectly accepted.
• Integers beyond binary64's exact range can be rounded before comparison. For example, 9007199254740993 no longer matches itself.

DAPO-Math-17K defines its labels as integers, so the scorer should validate that contract instead of truncating or rounding labels.

Change

• Parse normalized labels with Decimal rather than binary floating point.
• Reject malformed, non-finite, and non-integral labels with a contextual ValueError.
• Canonicalize valid integral decimal and exponent spellings without losing large-integer precision.
• Add CPU regressions for the false positive, precision loss, accepted integral spellings, and rejected label classes.

General symbolic verification and the rm_type=math path are intentionally unchanged.

Verification

• Base regression proof against origin/main: failed because 0.5 was accepted as 0 and 9007199254740993 was rounded.
• PYTHONPATH=. /tmp/slime-dapo-venv/bin/python -m pytest -q tests/test_rm_math_dapo.py — 34 passed.
• ruff check slime/rollout/rm_hub/math_dapo_utils.py tests/test_rm_math_dapo.py — passed.
• uvx --from black==24.3.0 --with click==8.0.2 black --check slime/rollout/rm_hub/math_dapo_utils.py tests/test_rm_math_dapo.py — passed.
• Codex autoreview — no accepted or actionable findings.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.