op4 HOSB: blanked-grid scoring primitives (stage 2)

[bitzic]

• host_reduce: BlankedCell, NonCausalModelError, reduce_blanked_nlls — HOST verdict over a (M,L) blanked-grid NLL array with two-filler + wrong-target witnesses; STRATUM-WEIGHTED mean (head/tail by true size) so forced tail coverage does not bias val_bpb vs compute_val_bpb
• val_bpb: build_blanked_grid (real-prefix-then-filler rows, one host-secret scored position each; stratified head/tail sampling) + per_position_nlls_blanked
• look-ahead is useless BY CONSTRUCTION: the answer (input[t+1]) is filler at every scored position; a causal model's score is identical to single-pass
• wrong-target witness is FRACTION-based (reject only if most wrong cells are sub-floor) so a confident honest model isn't false-rejected on a collision
• empty/degenerate grid + bad fractions fail loud (ValueError), never inf
• pure additive library, NO live-path wiring yet (compute_val_bpb / op4 unchanged) — stage 3 threads block-hash entropy + shadow->fail-closed
• 15 CPU tests: answer absent, honest==single-pass, look-ahead gains nothing, stratum-weighted unbiasedness, both witnesses incl collision tolerance, tail