intent

const DefaultWeight = 0.05

DefaultWeight scales the whole intent score before it is added to the ranking score.

DECISION(2026-08): the raw intent score reaches ~+4 (coverage 1.4 + constructor 2.0 + language 1.8) while the cross-encoder it is added to emits a 0..1 sigmoid — so an unscaled "prior" outranks the calibrated model by 4x and becomes the primary ranker. Measured consequence: on the gen corpus (queries written in code vocabulary, which the coverage term rewards) it looked like +0.05 Hit@1, but across 20 repos of docstring-prose queries it cost -0.24 Hit@1 — helping only on the corpus it was tuned on. Scaled to a tie-breaker it can still order near-ties without overriding the cross-encoder. Swept on both corpora (gen Hit@1/Hit@3, mean self-retrieval Hit@1 over 20 repos): 1.0 -> 0.65/0.87, 0.667; 0.15 -> 0.65/0.90, 0.839; 0.05 -> 0.64/0.90, 0.880; 0 -> 0.60/0.88, 0.904. 0.05 costs ONE gen case (inside the +/-1 noise floor) and buys ~100 rank-1s across the polyglot suite; going to 0 loses 4 gen cases, so the ranker does earn its place - it just must not outvote the model. REVISIT IF: the reranker changes scale (a different model or a logit output instead of a sigmoid).