fix: re-check the terminal size on every render pass [patch]

[matt-edmondson]

Fixes #111

The problem

UIApplication.Render() took the terminal size once, on the first pass that found the root element unsized, and never looked at it again:

if (RootElement.Dimensions.IsEmpty) { RootElement.Dimensions = ConsoleProvider.Dimensions; }

Resizing the window left every layout pinned to the size at launch — content drawn past the edge on a shrink, dead space around the UI on a grow.

The fix

Three parts, and the first alone would have been correct and never reached.

1. Compare the size on every pass. Render() now checks ConsoleProvider.Dimensions against the size the current layout was computed for, and adopts a new one when it differs. Confirmed SpectreConsoleProvider.Dimensions is live rather than cached: Spectre's Profile.Width has a null _width backing field unless a host sets one, so it delegates to AnsiConsoleOutput.Width on each read.

2. Re-arrange the whole tree, not just the root. IUIContainer.ArrangeChildren() sizes and positions its own children but does not reach theirs, and assigning Dimensions only invalidates — it does not re-run layout. So arranging the root alone relayouts the top level and leaves everything beneath it at the old size. ArrangeTree walks the tree parents-first, since a child container can only lay its own children out once it knows its new size. (UIApplicationTests.CreateLayout already works around the missing walk with a manual panel.ArrangeChildren().)

3. Wake the input loop on a timer. A resize delivers no input, so a loop parked in Console.ReadKey never reaches another render pass — following the triage note, this takes the portable polling option rather than SIGWINCH, which the console abstraction does not expose. Each tick compares the size and renders only when it changed, so an idle terminal still draws nothing; ResizePollInterval defaults to 100ms and is internal so tests can shorten it. The pending read is carried across ticks rather than restarted, which would otherwise leave several reads racing for the next key, and cancellation still ends a run blocked on the keyboard — the behaviour #117 added.

Between resizes, a size the host assigned to the root is left alone, as before. A resize overrides it, since the old size no longer fits the window.

Tests

TUI.Test/UIApplicationResizeTests.cs, six tests: a pass after a resize adopts the new size; it re-lays out a nested container and its leaf (a label wider than the resized terminal, so the width it lands at reports which size the layout used); a host-assigned root size survives a pass with no resize; a running application picks a resize up with no key pressed; polling does not redraw while the size is unchanged; and the resize is logged.

BlockingConsoleProvider.Dimensions is now lock-guarded and counts Clear() calls, since a resize test writes the size from the test thread while the loop reads it. RecordingLogger moved out of UIApplicationLifecycleTests into its own file, matching how the other test doubles are laid out, so both classes can use it.

Verified the tests fail without the fix, reverting each mechanism separately so each test is tied to the part it covers:

Reverted	Failing
A — one-shot sizing restored	RenderAfterAResizeGivesTheRootTheNewTerminalSize, RenderAfterAResizeRelaysOutEveryContainerBeneathTheRoot, AResizeWhileWaitingForInputRedrawsWithoutAKeypress
B — ArrangeTree made non-recursive	RenderAfterAResizeRelaysOutEveryContainerBeneathTheRoot, AResizeWhileWaitingForInputRedrawsWithoutAKeypress
C — resize poll removed from the input loop	AResizeWhileWaitingForInputRedrawsWithoutAKeypress

In each case the remaining tests still passed, including WaitingForInputWithoutAResizeDoesNotRedraw — correct, since none of the three reverts makes the loop draw on a timer.

Full suite: 154/154 pass. dotnet build TUI.sln is clean in Debug and Release across net10.0;net9.0;net8.0, 0 warnings.

One thing not verified here: driving TUI.App under a real terminal and resizing it. As #117 found, SetCursorVisibility issues a cursor-position query (ESC[6n) that nothing in a headless container answers, so the app never gets past hiding the cursor. The acceptance criterion is covered by the tests above instead.

Sequencing

#111's triage put this behind #109, which #113 closed. This branch is cut from main at b08abc6, which carries both #113 and #117, and touches Render() and ProcessInputAsync() on top of them rather than around them.

Docs

CLAUDE.md's rendering flow gains the resize pass and the note that ArrangeChildren() reaches one level only — the fact that made part 2 necessary.

🤖 Generated with Claude Code

https://claude.ai/code/session_01KmfzCZGV3tASUEeUcsXMNN

---

Generated by Claude Code

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud