[Marc-André Moreau (mamoreau-devolutions)]

Summary

The builtin engine now beats raw libghostty-vt on all 13 vt benchmark workloads, up from 8 of 13. The five that were behind or tied are now ahead:

Workload	Before (#61)	After
cjk	0.77×	1.12×
emoji	0.80×	1.21×
utf8	0.77×	1.36×
cursor-motion	0.90×	1.09×
scroll-region	0.96×	1.11×

Ratios are builtin speed ÷ Ghostty speed in the same run, on one Windows x64 laptop. Screens are identical to Ghostty's on every workload except styled, whose known one-row difference is unchanged. Full table and methodology: docs/performance-benchmarks.md.

This PR also fixes windows ending up larger than the screen, or off it, after a display is disconnected.

One commit per change, in order. Each perf commit builds and passes the Core tests on its own.

Builtin engine

• PrintRunes append loop: the loop moves into a small static method. Each rune is one OR into precomputed template bits and one 16-byte store, with widths from a table where ambiguous widths are already resolved.
• Short prints (cursor-motion, yes): the print templates are cached until attributes, protection or shell-integration state change. A single printable byte is one store. The DEC graphics check is inline.
• CSI dispatch: parameters are parsed straight into the arrays the engine receives, and the parser calls the sealed engine directly. Plain sequences skip the special cases and go straight to the switch on the final byte.
• Combining text as interned IDs: one CombiningTable per buffer. Lines store int IDs in an array that is reused when the line is recycled. A cluster seen before allocates nothing, and the table compacts so it cannot grow without bound. Plain combining marks and ZWJ skip the per-character path. utf8 allocation fell from 9.4 MB to under 0.1 MB per MiB.
• Vectorized UTF-8 decode: runs of 3- or 4-byte sequences are decoded four at a time from 16-byte Vector128 loads, so Arm64 gets it too. They follow the scalar decoder's validity rules, and anything else falls back to it. CJK decoding is about 3× faster.

Every new path is checked against the per-character or byte-by-byte path by differential tests. Each check was also deliberately broken to confirm a test fails.

Benchmarks

• Invoke-TerminalBenchmarks.ps1 -MinMs passes --min-ms to the bench. This laptop's CPU leaves turbo about 2.7 s into sustained load. With 500 ms samples, runs landed on both sides of that drop and spreads reached 50–180%. 3 s samples keep them around 5–20%.
• docs/performance-benchmarks.md has the round 5 results, the engine changes, new per-stage costs and what remains. It also records that prefetching recycled scrollback lines was tried and measured slower.

Window placement

Windows moves windows off a disconnected display without shrinking them, so the terminal could overhang the laptop screen with its caption buttons out of reach. A saved position could also point at a display that is gone.

WindowPlacement.Fit leaves a window alone when the displays cover all of it, so a window spanning two connected displays stays put. Otherwise it shrinks the window and moves it onto the display it overlaps most, or the nearest one. MainWindow applies it:

• at launch, when a saved or requested position is restored, before maximizing;
• on Screens.Changed;
• when the window returns to Normal.

A maximized or full-screen window keeps its state; its saved normal bounds are fitted instead. Windows Terminal only checks the top-left corner at launch and never shrinks.

Testing

• Core tests: 1,191 passed.
• App tests: 632 passed, 9 of them new for WindowPlacement.
• Headless UI tests: 670 passed.
• -Baseline gate passed after each perf change.
• The window fix has not been tried with a real display disconnect.

🤖 Generated with Claude Code