recall: evaluate recency-weighted cluster centroid for candidate_l2s nomination

[toejough]

Deferred from recall v2 (eval-gated follow-up, Joe's call).

Recall v2 already applies recency twice: in the matched-set ranking (recency-biased cosine decides what's in and on top) and via use-only activation (only notes the agent actually used get LastUsed bumped — so a superseded-but-surfaced note drops in rank as more recently-used notes are activated, and fades).

The idea: weight the vector sum of each cluster's note members by their recency scalars when computing the centroid used to nominate the within-cluster top-5 candidate_l2s (and the cluster representative):

centroid = Σ(recency_i · vec_i) / Σ recency_i

This reinforces drift toward recently-used notes.

Why deferred, not dropped:

• It's a third recency lever on the same axis — risk of over-rotation toward recency (the same risk the baseScore relevance floor was chosen to avoid).
• In v2's within-cluster nomination it rarely changes anything: the top-5 only re-orders when a cluster has >5 note members, which is rare (~dozens of notes across ≤7 clusters).
• It's an empirical question: use the recency-eval harness to test whether nomination over-surfaces superseded notes. Add the weighted centroid only if the eval shows it's needed, calibrated to that signal.

Note: candidate_l2s is the YAML field name (kept for back-compat); the tier vocabulary collapsed to "notes" in recall v2.