fix(delta): decoy displacement must exceed the event; add the missing BND decoy

[joshfactorial]

Two defects in the negative control, found while cross-checking job 20824321. Neither changes a caller result; both change whether the caller results are believable.

1. The decoy did not displace large events

It shifted every record a flat 2 Mb. Truvari matches sized variants on reciprocal overlap, so a 19.6 Mb deletion moved 2 Mb still overlaps ~90% — and matching it is correct behaviour.

Measured: decoy_scoreable recall=0.202, TP=17, and all 17 were larger than the shift:

n=17   min=2,048,336   median=4,255,928   max=19,619,570      (shift was 2,000,000)

Not one match was smaller than the displacement. The control was reporting "the shift is too small" while its warning text claimed "the scorer is matching far more loosely than intended" — it misdiagnosed its own failure, which is worse than staying silent, because it points the reader at the wrong subsystem.

Displacement is now max(2 Mb, 2 × |SVLEN|), so it always exceeds the event and overlap is zero by construction. END moves with POS, or the record would describe a span it no longer occupies.

Verified on hand-computed cases — shift exceeds span in every one, so no shifted interval overlaps its original:

event	shift	POS	span
500 kb DEL	2 Mb (floor)	1,000,000 → 3,000,000	preserved
19.6 Mb DEL	39.24 Mb	5,000,000 → 44,239,140	preserved
2.05 Mb DUP	4.10 Mb	3,000,000 → 7,096,672	preserved
BND (no SVLEN)	2 Mb	9,000,000 → 11,000,000	END == POS

2. decoy_truvari was never called for BND

It ran once, for scoreable. So the headline manta_BND = 0.935 had:

positive control	✅ selftest_BND = 1.000, all 62 records
negative control	❌ absent

That missing half is what distinguishes detection from coincidence. --bnddist sets a proximity window, and nothing measured how often a displaced breakend still lands inside it.

BND is also where the decoy works best: a breakend is a point (END == POS), so any shift displaces it completely — no size-scaling needed.