Fix KING_QUEEN_PROXIMITY_DANGER's missing material scaling and settle

on -14/count after isolated self-play testing.

The 2026-09-14 commit's -32/count rescale (and a -21/count trim of it)
both tested net negative in isolated self-play against a file-storm-
only baseline (2780 games score 0.490 llr -1.91; 520 games score 0.480
llr -0.63) despite the underlying regression signal looking directionally
real. Root cause wasn't the raw magnitude -- this term applied at full
strength regardless of attacker material, unlike the main counter-based
king-safety score just above it, which tapers via
REDUCED_MATERIAL_DOWN_SCALER/8. Adding that same scaling here was the
bigger fix (770 games score 0.494 llr -0.34, nearly flat); trimming the
now-scaled magnitude a further ~33% to -14/count tipped it positive
(690 games score 0.522 llr +0.60, trending toward H1).

KING_BACK_FLIGHT_BONUS (still +1/square, unchanged since 4ff9170) rode
along in all of these isolated matches without being implicated -- no
evidence yet that it's a problem on its own.

Co-Authored-By: Claude Sonnet 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01P9EDZVcqu6E2bNRm4zDrTf

0 files changed, 0 insertions, 0 deletions