Tags: kierank/x262

MPEG-2: improve b-adapt 2

--b-adapt 2 lost to --b-adapt 1 in ways unrelated to its placement
decisions

1. mbtree boundary-anchor over-boost. The keyint-capped trellis types
   the anchor right before the next keyframe as P before mbtree runs,
   and the transient all-P typing past the cap then chains the whole
   window's propagation into it -- a large QP boost for a frame whose
   quality the imminent keyframe replaces (mobile_cif crf12: QP 21.5 /
   2816 B vs b-adapt 1's QP 24.8 / 1376 B, ~1.4 KB wasted per 15-frame
   GOP). b-adapt 1 dodges this purely by ordering: its boundary-
   oblivious typing leaves that frame B while mbtree runs, and only the
   post-mbtree keyframe enforcement flips it to P, so it never receives
   propagation. Retype the pre-boundary anchor B around the mbtree call
   (decision-neutral; only qp offsets move).

2. Interlaced lowres distance-parity artifact. The frame-based lowres
   underprices even reference distances on combed content because field
   parity survives even frame distances while odd distances comb: on
   the src10 pan, lowres says P(d2)/P(d1) = 0.93 where real coded bits
   (CQP, forced IPPP/IBPBP/IBBP structures) say 1.10. That hands
   single-B alternation a fake advantage in the path sums and collapses
   pan GOPs to IBPBPBP. Real cost is non-decreasing in distance (field
   MC covers the parity; also verified 1.17 on a second interlaced
   clip), so clamp each anchor's path cost to its cost at shorter
   distances, interlaced only. Increase-only, so the path search's
   pruning threshold stays sound. Fades are the real exception (longer
   distances can genuinely be cheaper), so the clamp is vetoed where
   the lookahead weights analysis found brightness weights; dissolves
   are not detectable that way and keep a small cost on dissolve-heavy
   content (ntia tea: +0.36% vs unfixed), against -1.0% on the pan
   archetype.

BD-XPSNR of b-adapt 2 vs b-adapt 1, before -> after (100 frames,
--preset slow --bframes 2 --keyint 15, crf 6/12/18/24; BD-SSIM2 agrees
on every clip):

  foreman  CIF-p    -0.32 -> -0.39   (win improves)
  mobile   CIF-p    +1.12 -> +0.77
  src10    SD-576i  +1.91 -> +0.64
  src13    SD-486i  -0.54 -> -0.54   (win, output unchanged)
  park_joy HD-p     +0.49 -> +0.39   (+0.45 at 200 frames)

For now, b-adapt 2 still stays off slower presets. The residual is a
mbtree free-run interaction that favors b-adapt 1's transient typing
dynamics regardless of placement, not a placement mis-ranking. Forcing
b-adapt 2's placements under b-adapt 1 costs only +0.17% on mobile.