fix(reward): grade equation, scientific-notation, and empty-boxed math answers

[philluiz2323]

Summary

Closes #107

normalize_math_answer / _check_math scored several correct MATH-500 answers as wrong because the model wrote them in an equivalent but unnormalized form:

• �oxed{x=5} vs 5 — the variable assignment was never reduced to its value;
• �oxed{5 imes10^{3}} vs 5000 — after imes -> *, 5*10^{3} matched neither float() nor the fraction path, and the sympy fallback reads ^ as XOR;
• an empty �oxed{} was committed as the (blank) answer instead of falling back to the last number in the text.

Same class of false negative as the merged thousands-comma fix (#35).

What Changed

• src/trinity/orchestration/reward.py
  • normalize_math_answer: drop a leading single-variable assignment ^[a-z]\w*= (guarded by a (?=[^=]) lookahead and an identifier-only LHS, so inequalities like x<=5 and == are untouched); rewrite trailing scientific notation a*10^{b} -> aeb so it parses as a float.
  • _check_math: treat an empty/whitespace �oxed{} like a missing box and fall back to extract_last_number.
• tests/test_reward_math_forms.py: 22 offline cases — the fixes plus guards that wrong values, inequalities, and empty-with-no-number still score 0, and that existing forms (fractions, commas, negatives, tuples) are unchanged.

Why This Matters

These answer shapes are common on MATH-500, so the grader silently under-counted correct answers in both the training reward and reported eval accuracy. The fixes are guarded to avoid introducing false positives.

Verification

$ PYTHONPATH=src python -m pytest tests/test_reward_math_forms.py -q     # 22 passed
$ PYTHONPATH=src python -m pytest tests/ --ignore=tests/smoke -q          # 198 passed

Notes

A docs/JOURNAL.md entry was added per repository protocol (AGENTS.md §6). Unit-bearing answers (�oxed{5 ext{ cm}} vs 5) are deliberately left for a follow-up — stripping ext{...} risks false positives on textual answers like ext{Monday}.