Eval: rebuild the graph with #399's extraction and re-measure

[antejavor]

Part of #390. From the #399 grilling. Depends on #413.

What

Rebuild the eval graph once with the new extraction, then measure:

• graph quality (graph-quality) against the baseline taken on today's graph;
• one full-hybrid run on LongMemEval's judge as the no-regression check (today 84–85/100; single-session-assistant 10/11 is the one to watch, since assistant-turn edges shrink).

Acceptance

Both reports, compared with the baselines.

[antejavor]

Dropped with map #390's close: graph correctness, not retrieval, and recall already scores 89/100. The decisions stay in #399's comments; revive under #386 if extraction quality becomes the goal.