Follow-up to #1578: judge prompt placeholder injection ({gold} splice); unseeded bootstrap CI bounds; metric crash recorded as zero with SUCCESS status; allow_inclusion grades by substring #1697

Summary

Four defects at pin 3c52bd0 (extends #1578, which was fixed in #1588).

1. Judge prompt placeholder injection

Sequential str.replace over the whole prompt means a model answer containing the literal {gold} splices the reference answer into the predicted-answer section. A wrong answer scores 1.0.

2. Unseeded bootstrap confidence bounds

The arena bootstrap resamples with no seed; three fresh-process runs on identical battles produce different 97.5 percent bounds (0.6875, 0.7203, 0.7828).

3. Metric crash recorded as zero with SUCCESS status

The adapter's except-handler writes value=0 and keeps the default SUCCESS status, so a crashing metric plugin counts as a wrong answer in the mean denominator, contradicting the project's own ScoreStatus contract.

4. Accuracy with allow_inclusion grades by Python in

A list reference gives full credit for one-of-several required letters while the complete answer scores zero; alias lists become exact-match case-sensitive; string references become substring.

All four reproduce with deterministic offline stubs. Available as PRs with tests.