repaintColumns cleared and redrew only [start,end], but a neighbouring cell's
glyph can overhang into that span. The clearRect erased the overhang and the
neighbour was never redrawn, leaving black notches through the line 1-2 cells
before the cursor that survived until a full rerender.
Redraw text for a couple of extra cells on each side, clipped to the cleared
span, so overhang from just-outside cells is restored without touching their
own cell areas. Keeps the per-column repaint efficiency (vs the full-row
repaint debug toggle, which fixed the bars but repainted every dirty cell).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
-Djprototerm.fullRowRepaint=true (or JPROTOTERM_FULL_ROW_REPAINT=1) bypasses the
per-column repaint in renderChanged and repaints the whole row, to bisect the
stale black-bar artifact that appears near the cursor and survives until a full
rerender.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The snapshot bucket lumped ghostty's native dirty-state update together with
the Java-side cell marshaling. Time them separately to see which half of the
~7ms/frame snapshot cost (now the dominant frame cost after the detectShift
hoist) is the real target.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
rowFingerprint(row) is invariant across the delta loop but was recomputed for
every candidate delta, making shift detection O(rows^2 x cols) on large changes
(full-screen scroll). Precompute each changed row's hash once, dropping it to
O(rows x cols). Profiling showed fingerprint hashing at ~74% of frame time under
heavy scroll, dominated by this loop.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Gated behind -Djprototerm.profile=true (or JPROTOTERM_PROFILE=1), accumulates
per-frame nanos into snapshot/fingerprint/draw/frame-total buckets and dumps
to stderr every N renders. Splits the three suspected render costs: native
snapshot marshaling, fingerprint hashing, and canvas draw recording.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
