judging: reject unrenderable prompt templates at construction instead of aborting the run

[ChrisW09]

Fixes #15.

A custom prompt_template with literal braces passes check_template_placeholders (all required placeholders present) but explodes at score time:

prompt_template = 'Q: {question} A: {answer} S: {scale} Reply as JSON like {"grade": {grade}}'
# → KeyError: '"grade"' at the FIRST judge call

The KeyError escapes LLMJudge.score (rendering happens before the try, which only catches LLMError), propagates through the bare asyncio.gather in judge_results, and aborts the whole judging phase — after the answer phase already spent its tokens. Same for any typo'd placeholder ({questions}).

Fix:

• check_template_placeholders now trial-renders the template with dummy values for every supported placeholder and raises a ValueError explaining {{ }} escaping and listing the supported placeholders. Both documented template paths (Rubric.prompt_template, LLMJudge.prompt_template) call this at construction, so the failure surfaces at study-definition time.
• build_judge_prompt wraps its real .format the same way, so a template that bypassed validation produces one clear actionable error instead of a bare KeyError.
• The missing-placeholder check stays a warning (those templates still render).

Regression tests: JSON-example template rejected on both paths with an "escaped" hint, typo'd placeholder rejected, and a correctly {{ }}-escaped JSON template renders as expected.

Test: full cafe-core suite — 110 passed (1 pre-existing failure = #4, fix in PR #34).

🤖 Generated with Claude Code