Experimental (PoC): xBR/EPX brick filtering for the iso HD view

[noctonca]

Status: experimental PoC, Draft, not for merge. Stacked on #314 (the iso view-scale foundation); the base will retarget to main if/when that lands. Sharing it so the approach and one engine finding are on record.

What it is

At an iso zoom (iso scale > 1, from #314) the brick background is nearest-stretched and blocky. This filters that background with an edge-directed upscaler and composites it over the unscaled, crisp 3D models, so the floor and walls smooth out while characters and objects stay sharp.

Console toggle (interior, zoom > 1):

iso scale 2     # zoom first
iso epx 0       # off (nearest bricks)
iso epx 1       # EPX-blend (cheap, mild)
iso epx 2       # xBR (edge-directed)
iso epx 3       # xBR supersampled (SSAA, smoothest)

How it works

• Recover the native bricks by downscaling the zoomed Screen by the zoom factor (no second AffGrille pass, which corrupts shared render state and makes moving objects draw black).
• Filter to 2x native, then bilinear-resample to the display, so it works at any zoom > 1 (zoom 2 stays bit-identical to the direct path).
• Composite at present time, via a new scene-level present callback that runs under the console/touch overlays: brick pixels (where Log == Screen) take the filtered RGB layer; model pixels are re-converted from Log through PtrPal.

xBR and its YUV edge metric here are written from the published algorithm, not ported.

One finding worth recording

Going RGB-at-present surfaced a pre-existing engine quirk: the SDL present palette misses the 216-255 range in iso scenes (those entries are black, while PtrPal holds the real colours, e.g. the translucent ghost's blue). Anything using that range presents black even though it is correct in Log, and --screenshot hides it entirely because it saves PtrPal-converted Log. The composite works around it by re-converting model pixels through PtrPal. Full write-up in docs/ISO_XBR_FILTERING.md.

Scope and status

• Experimental, off by default, interior + zoom > 1 only.
• This branch is intentionally unoptimised (per-camera-cut rebuild is slow). The perf work (skip-when-unchanged gate, luma-only metric, threaded filter, LRU layer cache: xBR 311 -> 61 ms, SSAA 1470 -> 189 ms at 1280x960) is preserved on feat/iso-xbr-optimized and documented for re-applying once the design settles.
• The camera snaps between fixed views with its own transition delay, so the rebuild lands inside an existing pause; steady-state is just the composite.

Not proposing this for merge.