feat(benchmark): harden PersonaMem AML evaluation - #37

Summary

• add a complete PersonaMem-v2 AML-local MCQ evaluation report
• make sharded PersonaMem answering honor OPENROUTER_ANSWER_MODEL, with the existing Qwen default as fallback
• make resumed answer sharding remove only stale generated input shards while preserving all persisted answer artifacts
• add strict shard-answer merging with duplicate, missing-ID, and unexpected-ID validation before MCQ scoring

Result

• completed 5,000 / 5,000 questions
• official MCQ accuracy: 1,945 / 5,000 (38.90%)
• 60 empty provider answers (1.20%)
• document that this AML-local result is not leaderboard-comparable and cannot claim Recall/Hit without QA-level source IDs

Validation

• python -m py_compile benchmark/aml-local/reshard_answers.py benchmark/aml-local/merge_answers.py
• merged and validated 5,000 unique answer IDs before scoring
• ran vendored PersonaMem MCQ exact scoring for all 5,000 records
• git diff --check