fix(renderer): repaint unchanged rows inside hard-scrolled regions

[dmccaffery]

After a successful hardware scroll, scrolln shifted the model of the physical screen on curbuf only, while the repaint loop in Render skips lines by newbuf.Touched. A row inside the scrolled region whose newbuf content is identical to the previous frame at the same index was therefore never repainted and kept displaying the scrolled-in content, leaving stale duplicate rows on screen.

Touch the scrolled region on newbuf as well, so every row whose physical cells moved is re-diffed against curbuf. Rows that already match emit nothing, so the optimization stays intact: benchmarks show no measurable regression (sec/op and bytes/frame unchanged within noise on Apple M2, benchstat n=10).

Adds a regression test asserting that curbuf matches the rendered frame after every render, plus scroll-path benchmarks for the unique- and repeated-row cases.

Benchmark results — issue #137 fix (touchLine(newbuf, …) in scrolln)

Renderer hard-scroll benchmarks (BenchmarkRendererScroll*, new in terminal_renderer_hardscroll_test.go):
an 80×24 fullscreen viewport with scroll optimization enabled, scrolled up one line per frame.
Compared with benchstat, n=10 runs each. goos: darwin, goarch: arm64, cpu: Apple M2.

Time per frame (sec/op)

Benchmark	Before	After	Delta	p-value
RendererScrollUniqueRows	92.82µ ± 3%	91.41µ ± 3%	−1.51%	p=0.035
RendererScrollRepeatedRows	97.67µ ± 1%	97.63µ ± 1%	~	p=1.000 (n.s.)
geomean	95.21µ	94.47µ	−0.78%

Bytes written to the terminal (bytes/frame)

Benchmark	Before	After	Delta	p-value
RendererScrollUniqueRows	56.27 ± 0%	56.26 ± 0%	~	p=0.650 (n.s.)
RendererScrollRepeatedRows	97.22 ± 0%	96.05 ± 0%	−1.20%	p=0.000
geomean	73.96	73.51	−0.61%

Allocations (B/op)

Benchmark	Before	After	Delta	p-value
RendererScrollUniqueRows	9.562Ki ± 0%	9.562Ki ± 0%	~	p=1.000 (n.s.)
RendererScrollRepeatedRows	6.234Ki ± 0%	6.216Ki ± 0%	−0.28%	p=0.000
geomean	7.721Ki	7.710Ki	−0.14%

Allocations (allocs/op)

Benchmark	Before	After	Delta	p-value
RendererScrollUniqueRows	1.405k ± 0%	1.405k ± 0%	~	p=1.000 (n.s.)
RendererScrollRepeatedRows	457.0 ± 0%	453.5 ± 0%	−0.77%	p=0.000
geomean	801.3	798.2	−0.38%

Summary

No performance regression from the fix. Touched-but-unchanged rows re-diff against curbuf and emit
nothing, so time per frame is unchanged within noise. The repeated-rows case (the bug scenario) writes
slightly fewer bytes after the fix because keeping curbuf accurate avoids spurious repaints on
later frames. Correctness: the issue's fuzz harness went from 5/38 corrupted trials to 0/4143.

Fixes: #137

• I have read CONTRIBUTING.md.
• I have created a discussion that was approved by a maintainer (for new features).