[GUI] Fix truncated Help, Cancel and OK buttons

[arnodelorme]

The bottom buttons were sized in normalized units, so their pixel width followed the figure scaling factor and their labels were clipped in narrow dialogs (28 px wide buttons on a 3440 px display, 56 px on a 560 px default figure). They are now sized in pixels from the rendered label, and the figure is widened only when the row does not fit.

Checked on pop_resample and on the single help, double help, no help, wide, geomvert and geom layouts, plus a full pop_resample round trip on sample data.

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.

[arnodelorme]

Code review

• Overall assessment: Fixes the clipping in every layout tested; the wide-dialog margin regression raised in the first pass is now addressed in 7ab34de.
• Highest-risk area: adjustbuttons in functions/guifunc/inputgui.m.
• Merge recommendation: Safe to merge.

Layouts, develop vs this branch

MATLAB R2025b, macOS, 3440x1440 display. Six inputgui layouts rendered in 'mode','plot', once with develop's inputgui.m shadowed in and once with this branch.

Each pushbutton's pixel position(3) measured against its text extent(3):

layout	develop (px)	this branch (px)	text needs (px)
pop_resample / inputdlg2	27 / 27 / 27	80 / 80 / 80 (figure 256 -> 281)	39 / 55 / 25
no help button	27 / 27	80 / 80	55 / 25
two help buttons	27 / 27 / 27 / 27	84 / 96 / 80 / 80 (256 -> 389)	64 / 76 / 55 / 25
geomvert listbox dialog	23 / 23 / 23	80 / 80 / 80 (200 -> 277)	39 / 55 / 25
long okbut / cancel labels	22 / 22 / 22	80 / 144 / 160 (200 -> 421)	39 / 124 / 140
wide dialog (1774 px)	94 / 94 / 94	80 / 80 / 80	39 / 55 / 25

Every clipped label now fits, and no layout is clipped on either side.

Margin fix (7ab34de)

The first commit used a fixed margin = 10, which ignored the proportional g.borders supergui applies to everything else. On the 1774 px wide dialog that pulled Help from x = 88 to x = 10 and OK from 85 px off the right edge to 10 px, so the button row no longer lined up with the column above it. adjustbuttons now reads the margins supergui had already used, and falls back to 10 px only when they are smaller than that.

wide dialog	left margin (px)	right margin (px)
develop	88	85
first commit	10	10
with 7ab34de	88	85

Narrow dialogs are unchanged by the second commit beyond a few pixels (pop_resample: Help x 10 -> 14, figure 276 -> 281), because the two borders nearly coincide there.

Other checks

• 'addbuttons','off' skips adjustbuttons entirely.
• 'mode', <figure> re-entry (the pop_headplot pattern) does not re-run adjustbuttons, so the pixel layout from the first call survives.
• 'geom' path, minwidth (700 -> figure width 700), and a uilist reduced to {} all behave.
• Full GUI round trip on sample_data/eeglab_data.set: pop_resample(EEG) returns srate 128, pnts 30504, history EEG = pop_resample( EEG, 128);, eeg_checkset clean.

Blocking

None.

Important

None remaining. The wide-dialog margin finding from the first review pass is fixed in 7ab34de.

Nits

1. Manual resize re-clips the labels. adjustbuttons ends with set(butobj, 'units', 'normalized'), so the pixel widths only hold at the size the dialog opens at; halving the figure width by hand takes the buttons back to 40 px. Keeping 'pixels' would fix that but would unpin the right-aligned buttons from the right edge, so this is a genuine trade-off rather than an oversight. Worth a one-line comment saying the normalized units are deliberate.
2. Figure widening clamps only the left edge. figpos(1) = max(0, figpos(1)-(rowwidth-figpos(3))/2) keeps the dialog on screen on the left but can push the right edge past the screen when rowwidth is large on a small display. Clamping against get(0,'screensize') on the right too would cost one line.
3. curext(3)+20 and minwidth = 80 are magic numbers; the 80 duplicates the 'width' 80 already given to each button at lines 182-193. A shared local would keep them in sync.

Test gaps

No automated check, which matches the repo (no central test suite). The regression is cheap to pin down if you want a snippet in the PR body or in AGENTS.md:

inputgui('geometry', {[2 1]}, 'uilist', { {'style' 'text' 'string' 'Rate'} {'style' 'edit' 'string' '128'} }, ...
         'helpcom', 'disp(1)', 'mode', 'plot');
h = findobj(gcf, 'style', 'pushbutton'); set(h, 'units', 'pixels');
p = get(h, 'position'); e = get(h, 'extent');
assert(all(cellfun(@(a,b) a(3) >= b(3), p, e)), 'button labels are clipped');

EEGLAB notes

1. allobj(end-3:end) and alignleft = [ true true false false ] are tied to the four g.uilist entries appended at inputgui.m:180-193 (help / help-or-empty / cancel / ok). That coupling is correct today but invisible from adjustbuttons. A comment at line 193 ("the last four uilist entries are consumed by adjustbuttons") would keep the two in sync.
2. The isbutton guard handles the {} slots, the nan values supergui writes for panel/uitable/link2lines, and the empty-cell case, so the no-help and single-help paths degrade cleanly. marginleft falls back to marginright when there is no left-aligned button, which is what keeps the no-help dialog symmetric (Cancel/OK unchanged at an 11 px right margin).
3. Not exercised: Octave. get(uicontrol,'extent') on a pushbutton is the only property used here that is not uniformly implemented across Octave versions; worth a quick check if Octave support still matters for guifunc.