cocoa: convert only the damaged rows, so reporting damage does something

[tannevaled]

The defect

Incremental present is wired end to end on macOS — Surface.Damage →
RenderDamaged → presentRects → setNeedsDisplayRect — and it saved
nothing. Measured on a window presenting at 60 Hz, same binary, one bool
changed:

	cost
whole-surface present	26.2% of a core
damage reporting	26.8%

Sampling the process said why. -drawRect: wrapped the whole framebuffer
in a fresh NSBitmapImageRep and drew it once, relying on AppKit's clip to
limit what reached the screen — which it does. What it does not limit is the
conversion:

-[NSView displayIfNeeded]
  -> drawRect:
     -> -[NSImageRep drawInRect:fromRect:...]
        -> -[NSBitmapImageRep CGImage]

A rep has no backing CGImage until it is drawn, so every present converted
every row whatever the invalid region said.

The change

A run of rows is contiguous in the framebuffer, so it can be wrapped as a
bitmap of its own. presentRects hands its rectangles to -drawRect: through
the window, DrawBands reduces them to merged row runs, and each run is drawn
as its own rep into its own destination rect. The conversion is then
proportional to what changed.

An AppKit-initiated draw — an expose, a resize, the first display — finds no
pending rectangles and falls back to one band covering the buffer, which is
exactly the draw this replaces.

What it is worth

Measured rather than implied: a window repainting a small region every frame
at 60 Hz costs 22.2% of a core before and 19.1% after, two runs each,
consistent in direction. That is ~14% off the frame cost — real and
reproducible, and well short of what the profile suggested. The conversion was
about a seventh of the cost, not the bulk of it; the rest sits below this layer
and is not addressed here.

Tests

The row arithmetic is in mapping.go, the OS-independent half the Linux
coverage lane measures, and is covered by construction:

• bands merge when they touch as well as when they overlap — converting a
  seam twice costs more than the row it saves;
• they clamp to the buffer and drop what falls outside;
• every row a caller asks for ends up inside a band, whatever the rectangles —
  a row left out is a row the screen keeps stale;
• BandDest places a band where the whole-buffer draw would have put those
  same rows, and adjacent bands meet exactly, since a gap between two of
  them is a stale line across the window.

internal/cocoa mapping.go stays at 100% statement coverage. The live
darwin repaint proof passes.

🤖 Generated with Claude Code