BM25 k1 sweep + I(f) perturbation sensitivity

[nikolay-e]

Context

Two recorded-but-open sensitivity checks from v2: BM25 k1=2.5 sits above the classical range on a stated hypothesis (short identifier docs, late saturation) with no sweep; I(f) importance priors (0.10/0.15/1.0) are uncalibrated domain guesses.

Tasks

• k1 sweep {1.2, 1.6, 2.0, 2.5, 3.0} x b {0.5, 0.75} on internal BM25 + RRF branch
• I(f) perturbation +/-25% and +/-50% on impact-need scoring
• Report deltas with CIs

Acceptance

• both closed as tables in paper v3; defaults updated only on CI-excluding-zero improvements