[tomlm]

VT DOOM (https://github.com/j4james/vtdoom) showed a black screen in Avalonia.Terminal and works in Windows Terminal. Diagnosed by running it headless through Porta.Pty into XTerm.NET and into a headless TerminalView, then at the user's real window size.

What was wrong, in the order found

1. Decoder crash → black screen. Growing the canvas for a taller band doubled the width too, overshot MaxSixelPixels, and the retry at "exactly the size asked for" took the new band's narrower width; copying the old rows ran off the end. The exception escaped Terminal.Write and ended the host's read loop, so ConPTY blocked and VT DOOM stopped on its first frame. Only the short dimension doubles now, and the retry never cuts into drawn pixels.
2. Aspect ratio ignored → picture a fifth of its height. VT DOOM scales vertically with the raster attribute's Pan;Pad ("5;1). Each sixel row is now Pan/Pad pixels tall (capped at 16). P1 stays ignored, as in xterm.
3. Private colour registers → a scatter of colours on black. VT DOOM sends its palette only when it changes (29 frames in 496). Registers are now shared by default, as on a VT340 and in Windows Terminal; CSI ? 1070 h still selects private registers, DECRQM reports it, and RIS now also resets the shared registers.
4. Dropped frames → freezes, and a permanent freeze after a key press. The doubled canvas (4,096 wide for a 2,188-wide frame) was what got tested against the budget. At 307×75 cells such frames were discarded — data arriving, nothing drawn; once the menu appeared, every frame. The budget retry now shrinks to the drawn extent.
5. GC pressure → input arriving in bursts. ~25 full collections a second while VT DOOM ran:
   • the decoder allocated a canvas per picture, per growth, plus an exact copy → now a per-terminal SixelCanvas with two reused buffers, in-place compaction, starting at the last picture's size;
   • SixelStrip held a row in one array past the large-object threshold, and a host capturing frames made every repaint copy it → strips are now blocks of 16 cells, allocated on paint and copied on write per block.

Measured: full collections 373 → a handful in 15 s; VT DOOM at 307×75 places every frame (150/151, the last one in flight) through Escape and Enter.

API change (3.0 is at rc001)

SixelStrip.Pixels is replaced by CopyPixelsTo(Span<byte> destination, int destinationStride, int columns), which hands out an ordinary contiguous BGRA image. The block storage behind it is internal; nothing outside depends on it. README's 3.0 section and drawing guide updated. Avalonia.Terminal's matching change is in its own PR.

Behaviour changes worth noting for 3.0: shared colour registers by default; the aspect ratio is honoured.

Tests

2,312 pass. New: the growth crash (fails without the fix), a picture inside the budget surviving the canvas doubling past it, aspect ratio (3:1, square == none, capped), shared/private/RIS colour registers, a reused canvas not showing the previous picture. Strip tests read pixels through CopyPixelsTo; the budget test works from real strip sizes.

🤖 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	2.74 → 2.72	-0.8%	±1%	4%
sgr-churn	0.00 → 0.00	0.00 → 0.00	7.26 → 7.33	+0.9%	±3%	9%
truecolor	0.00 → 0.00	0.00 → 0.00	7.70 → 7.62	-1.0%	±2%	6%
alt-redraw	0.00 → 0.00	0.00 → 0.00	11.53 → 11.71	+1.6%	±1%	4%
unicode	7.66 → 7.66	0.45 → 0.45	28.08 → 28.07	-0.0%	±3%	9%
flood	0.00 → 0.00	0.00 → 0.00	82.61 → 82.72	+0.1%	±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 3.0.0.0 mvid:8a972284-ba2c-4595-bf7a-75a1ded0c392
• head: XTerm.NET 3.0.0.0 mvid:80cc7062-6296-4011-878d-4ee399031774

---

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	2.72 → 2.72	+0.0%	±1%	5%
sgr-churn	0.00 → 0.00	0.00 → 0.00	7.25 → 7.33	+1.1%	±3%	9%
truecolor	0.00 → 0.00	0.00 → 0.00	7.62 → 7.62	+0.1%	±0%	5%
alt-redraw	0.00 → 0.00	0.00 → 0.00	11.54 → 11.71	+1.5%	±1%	5%
unicode	7.66 → 7.66	0.45 → 0.45	27.94 → 28.07	+0.4%	±1%	5%
flood	0.00 → 0.00	0.00 → 0.00	82.15 → 82.72	+0.7%	±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:80cc7062-6296-4011-878d-4ee399031774