sessions-graph: ANN vector search for recall on large histories

[antejavor]

Part of #390. Follow-up from #417.

Problem

Recall computes similarity exactly, over the asking user's own vectors (vector_search.cosine_similarity in Cypher), because vector_search.search has no filter. The cost grows linearly with one user's history. Measured on a synthetic user:

Messages	Recall, median
183 (largest eval user)	0.6 s
10k	0.7 s
100k	6.7 s

At 100k messages, the turns lane takes 3.0 s, the text lane 1.8 s (1.1 s of that is fetching the user's own turn ids), and each graph lane about 1 s.

What to build

The fallback agreed in #393: global ANN vector indexes (node and edge) with an over-fetch, filtered to the user afterwards.

• Measure the over-fetch needed to keep the exact top-k on a multi-user graph, and fall back to the exact scan when the filtered result comes up short.
• Use the ANN path only above a history size where the exact scan gets slow.
• The text lane's own-turn list is the other half of the 100k cost: filter by ownership in the traversal instead of passing a list.

Done when

• Recall at 100k messages is under 1 s.
• The parity run still holds (official judge within noise of g417-recall-r1, evidence recall not lower).

[antejavor]

Dropped with map #390's close. Recall's exact scan takes 0.7 s at 10k messages; reopen if a real user's history approaches 100k.