Eval: one user per question, retrieval scoped to it

[antejavor]

Part of #390

Problem

inject.py gives every injected session EVAL_USER_ID = "longmemeval-user", and hybrid retrieval searches the whole graph for every question. In LongMemEval each question's haystack belongs to a different person. In the #391 run (100 questions, 501 sessions):

• only 31% of retrieved turns and 17% of retrieved facts came from the question's own sessions;
• answers counted other personas' items: 4 weddings for 3, 6 furniture items for 4, another persona's $1,200 bag added to the right $800;
• one User holds 897 purchased and 291 lives_in edges, so the user_facts lane gathers 100 people's facts;
• cross-persona rows crowd the question's own evidence out of the top results. That drives most of the 25 retrieval and partial-retrieval misses.

What to do

• Each question gets its own User; its sessions hang off that user.
• Every hybrid lane is scoped to the asking user's sessions.
• If two questions share a haystack session, inject it under each question's user, or link the shared session to both users. Decide in the PR.
• Land it in the same eval-graph rebuild as Write turn links and edge source sentences at extraction time #392.

Acceptance

Retrieved context comes only from the question's own sessions. The baseline is re-measured on the rebuilt graph with one full-hybrid run.