daily-fix: parse_judge_json drop-never-coerce (eval/utils.py)

Overview / Motivation

Auto-filed by the /daily 2026-07-01 backfill problem sweep (route 2 — behavior/logic change requiring independent review).

Goal

Port drop-never-coerce judge-return handling (llm-judging rule 9, the #766 eval/belief.py precedent) to the shared binary-judge parser parse_judge_json in src/explore_persona_space/eval/utils.py and audit its call sites, so malformed judge returns are DROPPED (NaN/None excluded from aggregates) instead of silently coerced to a caller-supplied default.

Bug observed

• On #778 (2026-07-01), the capture-path judge logged "Failed to parse judge JSON, using default" for ~4,533 of the first 14,993 judge calls (~30% early; final tally ~8,225/87,055 ≈ 9%) — every one coerced to a default score instead of dropped (session a37975ac, 12:51Z).
• On #763 the same day, the run log showed repeated "[WARNING] Failed to parse judge JSON, using default" from src/explore_persona_space/eval/utils.py:20 on the BINARY judge path (session 653baadd, 06:12Z); the graded module has its own drop-not-coerce parser, the shared binary parser does not.

Why this needs fixing

llm-judging rule 9 (.claude/rules/llm-judging.md): a malformed / REFUSAL / out-of-range judge return carries no information and must be dropped from BOTH arms, never coerced — coercion biases the mean toward whichever arm the failures land in. eval/belief.py was fixed for exactly this in #766 (parse failures → math.nan, excluded by the caller's isnan filter); parse_judge_json still returns the caller's default dict on failure.

Proposed change (refine in planning)

• parse_judge_json returns None (or NaN-marked) on parse failure; per-call-site audit decides drop vs explicit caller-side handling; report per-arm dropped counts where aggregates are computed.
• Root-cause the ~30% parse-failure burst on the #778 capture path (likely truncated / non-JSON judge text) — a parse-rate that high also hints the judge prompt or max_tokens needs fixing.
• Raw judge text is saved on these paths, so historical scores are re-parseable offline; no data loss.

Scope / surfaces

• Primary: src/explore_persona_space/eval/utils.py (parse_judge_json) + its call sites (grep parse_judge_json across src/ and scripts/).
• Sibling precedent: the #766 belief.py fix; .claude/rules/llm-judging.md rule 9.

Constraints / invariants

• Do not change the graded-judge module (already drop-not-coerce).
• Dropped counts must be reported, never silently swallowed (fail-fast discipline).
• Ruff + relevant tests pass; add a regression test for the drop behavior.