[Enhancement] Automated regression eval for summary quality (gold-standard comparison)

[perrycan]

Right now the summary rules are tuned by manually comparing the manager's hand-written summary against the model output, day by day (see docs/SUMMARY_RULES.md).

It would be great to turn that into an offline eval script:

• Use a set of roll-call text -> expected summary samples as the gold standard;
• Run the model and score it structurally (category accuracy, missing/extra items, correct merge/split);
• Wire it into CI so rule changes are regression-tested — avoid "fix one rule, regress another".

Discussion on scoring metrics welcome.