fix(reward): extract \boxed{} before grading RLPR WebInstruct math (#170)

[minion1227]

Fixes #170

Problem

_check_rlpr_webinstruct in src/trinity/orchestration/reward.py scores WebInstruct-verified RLPR items (WebInstruct-verified-val_Avg2). For a math answer it fed the raw candidate prose straight into math_equal / normalize_math_answer — unlike _check_math, which calls extract_boxed / extract_last_number first. Since normalize_math_answer doesn't strip \boxed{...}, a correctly boxed answer never matched the bare gold and scored 0.0 — even though format_hint("rlpr") explicitly instructs the worker to "use a boxed final answer for math items". The requested output format was the one graded wrong.

from trinity.orchestration import reward as R
ref = {"ground_truth": "15", "source": "WebInstruct-verified-val_Avg2"}
R.score_text("rlpr", r"The area is \boxed{15}.", ref)      # 0.0  <- correct answer, scored wrong
R.score_text("rlpr", "The answer is 15", ref)             # 1.0  (plain number happened to work)
R.score_text("math500", r"The area is \boxed{15}.", "15") # 1.0  (math path extracts the box)

Impact: every WebInstruct-verified RLPR math item answered in the requested \boxed{...} format is a false negative — understating accuracy on that benchmark and corrupting the per-benchmark reward signal.

Fix

Extract the answer from candidate (and gold) before the math compare, mirroring _check_math:

cand_math = extract_boxed(cand) or extract_last_number(cand) or cand
gold_math = extract_boxed(gold) or gold
if math_equal(cand_math, gold_math):
    return True
return normalize_math_answer(cand_math) == normalize_math_answer(gold_math)

The or cand fallback is load-bearing: when a candidate has neither a box nor a number (a free-text answer), cand_math is the whole candidate, so non-math text answers grade exactly as before — the behaviour only diverges when an answer is actually extracted.

Distinct from #116 / #122

Those cover the RLPR MMLU-Pro choice path (letters E–J). This is the WebInstruct math branch — a different code path and a different fix. The letter-choice branch here is untouched.

Tests

Adds tests/test_reward_rlpr_webinstruct_boxed.py (offline, pure stdlib — no network/GPU):

• the boxed repro now grades 1.0; a boxed value wins over a stray earlier number; a number-in-prose without a box is recovered; a boxed LaTeX fraction matches 1/2; a boxed gold is unwrapped
• guards: the plain-number, free-text ("Paris"), and letter-choice paths are unchanged; empty candidate/gold still False

5 of the new cases fail on the old code and pass on the fix; the 10 guard cases pass either way. Full root suite: 261 passed. ruff clean on the changed lines (the 3 pre-existing findings in reward.py are on unrelated lines).