context-graph-eval: answer with the question date, compute over listed evidence, turns first

[antejavor]

Part of #390. Closes #398. Stacked on #389 (base feat/eval-hybrid-retrieval); GitHub retargets it to main when #389 merges.

Why

On the rebuilt graph (#392 + #396) and the answer gate (#402), the full hybrid scored 60/100. Most misses with the evidence already in context were answering errors:

• No question date. The eval never passed it unless --question-date was set, and it wasn't, so "how many weeks ago" answers counted from an invented "today" (2024-06-13).
• Miscounting. 4 weddings for 3, 4 plants for 3, 152 game hours for 140; "two weeks" and "a week and a half" never summed.
• Recommendations refused. All 6 preference questions answered "not in memory" under the "answer only from the rows" rule.
• Turns read past. About 45 fact rows came before the turns, which are the answer store.

What changes

• The question date always reaches the answer prompt. --question-date and RunPlan.question_date are removed. LongMemEval supplies it with every question, and in a real session it is simply now.
• The shared answer prompt (retrieval.answer_prompt, used by every strategy):
  • relative dates inside a row ("today", "yesterday") are relative to that row's date;
  • for counts, totals and intervals, list each item with its date, count an item mentioned twice once, then compute;
  • when rows disagree, the most recent one is current;
  • recommendation questions get suggestions built from the user's stated preferences;
  • a question about something the rows never mention, or with a premise they don't show, gets "not in memory", not an answer to a related question.
• Hybrid puts turns before facts, both in time order.

Measured

One full-hybrid run per variant on the rebuilt graph (port 7735, 100 questions, 5 sessions each), Sonnet 4.5 judge, answer gate:

type	before	r1	r2 (this PR)
temporal	10/26	20	22
multi-session	12/26	16	15
knowledge-update	13/16	14	15
single-session-user	14/15	15	14
single-session-assistant	11/11	10	10
preference	0/6	0	0
abstention	7/8	5	7
total	60	75	76

r1 lacked the false-premise rule, and abstention fell to 5/8: told to list and compute, the model answered a nearby question. r2 adds that rule and holds 7/8. Between r1 and r2, 5 questions flipped each way at the threshold, so ±4 per type is run-to-run noise.

Preference stays 0/6, now because of the judge, not the answers. The answers recommend what the question asks for (Premiere Pro resources for a user asking about its advanced settings; MICCAI and ISBI for a medical-imaging researcher), but the Coverage rubric checks for facts and the expected output is a description of a good answer. A preference rubric is filed as #404, kept out of this PR so the prompt and the judge don't change together.

Tests

context-graph/eval: 263 passed, 5 skipped against a real Memgraph; ruff and ty clean. New: the runner always passes the question date to the answer prompt; hybrid rows are turns then facts, each in time order.

Update: recommendation questions are never refused (fbb5e0f)

Re-judging with #406's preference rubric showed r2's false-premise rule also caught recommendation questions: "recommend a conference" names no conference in memory, so two answers fell back to "not in memory". Recommendation questions now always recommend, from the preferences the rows show plus general knowledge; "not in memory" stays for factual questions.

One full-hybrid run of this PR with #406's rubric (same graph and judge, answer gate): 80/100. Temporal 22/26, knowledge-update 15/16, single-session-user 14/15, multi-session 13/26, single-session-assistant 10/11, preference 6/6, abstention 8/8.