Scroll a dashboard taller than the pane, with a scrollbar

[shayan-ys]

What and why

On a short pane, watch mode cut off every row past the bottom with … N more rows. Watch mode draws on the terminal's full-screen view, which has no scrollback, and drawing inline can't redraw a frame taller than the pane without piling stale copies into scrollback. So devdash now scrolls inside the app, the way less and htop do.

• up/down scroll one row. These are new built-in keys, keys.scroll_up and keys.scroll_down, and can be remapped or disabled like the other built-ins. An integration that binds up or down now fails config validation with the existing remap hint.
• The mouse wheel scrolls too. On the full-screen view, most terminals send arrow keys for the wheel. devdash also turns on alternate scroll mode (DEC 1007) while it runs, then restores the previous setting with XTSAVE/XTRESTORE. It writes these sequences only when stdout is a terminal, and restores the mode even if restoring stdin fails.
• A scrollbar in the rightmost column shows which part is on screen. When the content overflows, the dashboard and its integrations lay out one column narrower to make room for it.
• Each frame reads the scroll offset once and re-checks overflow after the narrower layout, so a key press or a data refresh between those steps can't misalign the scrollbar.
• The footer lists the scroll keys only when the content overflows. --once now renders without a viewport, so it prints every row (it used to stop at a 10,000-row height).

Checklist

• uv run pytest and uv run ruff check . pass (116 passed; ruff clean)
• A test covers new behavior that a plausible bug could break (test_a_dashboard_taller_than_the_pane_scrolls_with_its_keys_within_bounds: thumb position at both ends, scrolling stops at both ends, narrower layout width, --once prints every row)
• CHANGELOG.md has a line under Unreleased for a user-visible change
• The README image is regenerated if the pane looks different: not needed, because the screenshot is taller than its content and shows no scrollbar. Regenerating it shows a one-row drift that already exists on main, so it stays out of this PR.

Verification

• Ran the real app in a 70×20 pty, rendered with pyte. It starts with the thumb at the top; 30 wheel-style ESC O B presses scroll the content and move the thumb down. On quit, the output contains the 1007 save, set and restore sequences.
• Rendered panes 1 to 3 rows tall without errors; --once printed all 77 rows of a real dashboard.
• CI is green on ubuntu and macOS with Python 3.12, 3.13 and 3.14.

Review

• Review path: material, agent code review, because the new built-in keys change the public config contract. CodeRabbit is not installed on this repository. No optional reviews were run.
• Accepted and fixed in e0d047b: the scroll offset race between the key loop and the render thread; overflow decided from a different snapshot than the one drawn; mode 1007 setup outside the cleanup scope; escapes written to redirected stdout; --once cropping past 9,998 rows.
• Declined: the width feedback loop never settling. The narrower layout only adds wrapped rows, so content that overflows at width W also overflows at W−1, and the width settles after one change. Only an integration that prints fewer rows when given fewer columns could make it flip back and forth.
• Remaining non-blocking findings: none.

[shayan-ys]

@coderabbitai review