Hold Sixel as the pixels each line shows (3.0.0)

[tomlm]

Why

A Sixel is content: it paints the cells it lands on, a later Sixel paints over it, and printing over a cell replaces it. 2.x kept every Sixel as a picture — a TerminalImage plus one LinePlacement per row — and split it wherever something covered part of it. A program that repaints small regions (a cursor, a hover, a diffing framebuffer such as SixelBuffer) left a patchwork behind.

Measured in Avalonia.Terminal at 307×77 running SixelBuffer's ControlCatalog in Sixel mode:

	pictures on screen	paint per frame
2.0.7	13,551 images, 58 MB	593 ms
2.0.8 (#175)	2,606 images, 23 MB	110 ms
this PR	76 strips, 11 MB	0.6 ms

What changes

• SixelStrip (new, public, read-only from outside): the Sixel pixels one line shows, one cell tall, BGRA straight alpha, with CellWidth/CellHeight, PaintedColumns, HasPixelsAt(col) and a Version that moves after every change. BufferLine.Sixel exposes it.
• Painting: PlaceImage keeps its cursor, DECSDM, scroll and mode-8452 behaviour, but for Sixel it paints each row into the line's strip. Opaque pixels replace, unset pixels (background select 1) leave what was there, partial alpha (inline PNG) blends.
• Erasing: SetCell, SetSingleWidthRun and Fill clear the cells' pixels; a strip with nothing left is dropped.
• Moving: CopyCellsFrom moves pixels with their cells, so ICH/DCH, insert mode and margin scrolls no longer leave Sixel behind.
• Copies: Clone/CopyFrom share the strip until either side writes (copy-on-write), so frame capture costs a reference.
• Rescaling: a Sixel laid out against a new cell size resamples the strip first, so one strip has one scale.
• Budget: MaxImageBytes counts strips; over budget, scrollback strips go first, oldest line first; what is on screen is never taken.
• iTerm2 inline images are painted into strips the same way.
• Kitty is unchanged. Kitty position/z deletes no longer touch Sixel (they never should have).
• Terminal.SixelPlaced (internal) lets tests see the decoded picture, which the buffer no longer keeps.

Breaking (3.0.0)

Sixel no longer appears in line.Placements / line.Images; TryGetPlacementAt / TryGetImageAt find Kitty pictures only; PlacementKind.Sixel is never created (the value stays so comparisons compile). A 2.x renderer keeps compiling and silently draws no Sixel, hence the major version. README has an "Upgrading to 3.0" section with the renderer change.

Tests

2,303 pass. The Sixel placement and lifetime tests now assert pixels (a striped picture checks which band landed on which row); new tests cover ICH/DCH moving pixels, transparent pixels showing through, repainting 50 times without growth, copy-on-write, cell-size rescaling, and the budget rules.

The renderer side is in Avalonia.Terminal (separate PR), verified pixel-identical between its Skia and classic paths.

🤖 Generated with Claude Code

[github-actions]

Perf comparison — this change, against its base

3 run(s) of each side, alternating on one machine. Allocation is a count and is gated exactly. Time is a measurement, so its gate is derived from the spread this job just observed in itself rather than fixed in advance.

corpus	bytes/char	gen0/Mchar	ns/char	Δ time	noise	gate
scroll-ascii	0.00 → 0.00	0.00 → 0.00	3.40 → 3.46	+1.8%	±3%	8%
sgr-churn	0.00 → 0.00	0.00 → 0.00	9.10 → 9.06	-0.5%	±1%	4%
truecolor	0.00 → 0.00	0.00 → 0.00	10.25 → 9.98	-2.6%	±3%	8%
alt-redraw	0.00 → 0.00	0.00 → 0.00	14.50 → 14.54	+0.3%	±3%	10%
unicode	7.66 → 7.66	0.45 → 0.45	34.92 → 35.95	+2.9%	±3%	10%
flood	0.00 → 0.00	0.00 → 0.00	96.08 → 96.05	-0.0%	±1%	4%

Each corpus is gated at max(4%, 3 × its own noise). A wide noise column means this runner was busy and the timing half of the table should be read as advisory; the allocation half is exact either way.

assemblies measured

• base: XTerm.NET 2.0.8.0 mvid:4d610f86-7357-4b04-a33a-2422bab70f0b
• head: XTerm.NET 3.0.0.0 mvid:7d82d65d-e009-4cdd-9e95-353f6fd4194a

---

Perf comparison — cumulative, everything since 2.0.7

3 run(s) of each side, alternating on one machine. Allocation is a count and is gated exactly. Time is a measurement, so its gate is derived from the spread this job just observed in itself rather than fixed in advance.

corpus	bytes/char	gen0/Mchar	ns/char	Δ time	noise	gate
scroll-ascii	0.00 → 0.00	0.00 → 0.00	3.40 → 3.46	+1.8%	±3%	8%
sgr-churn	0.00 → 0.00	0.00 → 0.00	9.18 → 9.06	-1.3%	±1%	5%
truecolor	0.00 → 0.00	0.00 → 0.00	10.30 → 9.98	-3.1%	±2%	7%
alt-redraw	0.00 → 0.00	0.00 → 0.00	14.46 → 14.54	+0.5%	±2%	6%
unicode	7.66 → 7.66	0.45 → 0.45	35.90 → 35.95	+0.1%	±2%	6%
flood	0.00 → 0.00	0.00 → 0.00	96.33 → 96.05	-0.3%	±1%	5%

Each corpus is gated at max(5%, 3 × its own noise). A wide noise column means this runner was busy and the timing half of the table should be read as advisory; the allocation half is exact either way.

assemblies measured

• base: XTerm.NET 2.0.7.0 mvid:2609fe33-3671-4fb3-9e45-c3e43c231eb3
• head: XTerm.NET 3.0.0.0 mvid:7d82d65d-e009-4cdd-9e95-353f6fd4194a