v3 experiment matrix: 16 cells, budget curve, per-role table

[nikolay-e]

B=8000. Tune on dcbench-calib; report on dcbench-holdout + public 1500. One full run, after pre-registration.

#	cell	why
1	EGO (v5)	anchor
2	PPR (v5)	anchor
3	BM25-int (v5, with gate — prerequisite 2)	anchor; the gate changes exactly this cell
4–8	HK t ∈ {1,3,5,10,15}	C2 grid
9	union (v2-style)	fusion baseline
10	sequential BM25→EGO	C4
11	fused-sum (calibrated)	C1
12	fused-noisyor	C1
13	fused + DPP	C3
14	discovery-union + fused	C5
15	EGO @ v4 config on dcbench	bridge to v2
16	fused, seeds = change_magnitude	restored v1 signal

• Budget curve {8k, 16k, 32k, −1}: winner and EGO only.
• Appendix ablations on calib: DPP λ-mix; calibration isotonic vs Platt vs none.
• Per-cell artifacts as in v2 (per-instance CSVs, seeds, CIs); per-role table from dcbench annotations.
• Double-run determinism check and python -m eval equivalence on the stratified sample before the run.