[Tiaxi]

Closes #542.

While zooming, both plots briefly drew the background ("no data") where older data could have stood in until the new zoom's data arrived. Frame sequences recorded from a GL window under a headless KWin (each rendered frame read back) showed where the black came from.

Where the black came from

Spectrogram

• Missing columns drawn black. When the new zoom's first columns arrived, the canvas was rasterized again from the ring, and every column not decoded yet was filled with the background. A centered zoom out (and Reset Zoom, which is one) ends the zoom preview as soon as data comes, so the view collapsed to a sliver at the playhead and filled in over a few hundred milliseconds.
• Rolling history thrown away. A rolling zoom restart clears the history, and a rolling session never decodes it again. So the whole view went black and refilled from the right edge as playback went on.
• Zoom-out margins. The zoom preview only scales the canvas, so a zoom out showed the background beside it. The ring usually holds the columns there.

Waveform

• In the per-channel view (the default), the overview fallback draws nothing, because the overview is one envelope with no channel lanes. While playing, the view is drawn in playback tiles, which leave no cache, so a zoom out (and Reset Zoom) could not be deferred. It showed at once over the overview fallback, and the lanes went empty until the new detail arrived (about 110 ms, 8 frames in the recording).

What changed

Spectrogram (SpectrogramItem)

• The canvas is premultiplied ARGB. Missing columns are transparent instead of black, and the canvas textures carry alpha.
• Two layers are drawn under the canvas, each placed by the time of its columns (the mapping staleCanvasAlignmentLocked already uses), so they follow the preview, scrolling and playback:
  • Backdrop: the canvas that a canvas with missing columns replaced. It keeps the view before the zoom at full detail until the new zoom's data covers the view, then it is dropped.
  • Reserve: a worker rasterization of the ring's columns around the view. It covers the edges a zoom out reveals from its first frame. It is kept up to date as playback moves on; a zoom out draws it again at the zoom asked for; and it captures the ring just before a zoom's session clears it. Each new reserve is composited over the one before on the worker, so older data is never dropped.
• Both layers are dropped when their times stop being the view's: seeks, track changes, a mode change, a new projection (height, dB range, log scale), an epoch re-anchor, or released storage.
• No timers. The UI and render threads only place textures; the rasterizations and compositing run on the existing raster pool, with serials and a validity counter so stale results are discarded.

Waveform (WaveformEditorItem)

• A zoom out is held while playback tiles can still draw the old view, as it already was while a cache could. It shows once the wider view's detail arrives.
• Where the overview cannot stand in (the per-channel view), the detail the view has stands in for the part it reaches. Detail sparser than a pixel holds each point until the next one, instead of leaving gaps.

Docs: docs/SHORTCUTS.md describes what the plots now show while zooming.

Tests

• spectrogramZoomShowsOlderDataUntilTheNewArrives (centered and rolling): wheel steps out, data for them, a step in and back to 1. Each frame is checked through its scene-graph nodes (canvas, backdrop and reserve, with the canvas's alpha): no column that showed data shows the floor in the next frame, and the backdrop stands in wherever the canvas lacks columns. It fails without the reserve, and without the backdrop.
• waveformEditorPerChannelZoomShowsOlderDetailUntilTheNewArrives: in the per-channel view, a reset and a zoom in past the detail's points keep the wave in both lanes. It fails without the fix.
• waveformEditorZoomOutWhilePlayingHoldsTheTiledView: while playing in tiles, a zoom out is held with no empty lane columns, then shows when its detail arrives. It fails without the fix.
• ./scripts/run-tests.sh --ui-only: 102 of 102 passed, with no warnings. Rust is unchanged.

Frame sequences

Recorded under a headless KWin (OpenGL), with each frame of the plot read back while zooming by wheel steps and Reset Zoom, in rolling and centered mode, against the fixture library. Counts are frames in which columns that showed data turned to the background, before → after:

Plot, mode	Before	After
Spectrogram, centered	6 frames, 2500 columns	0
Spectrogram, rolling	18 frames, 1233 columns	2 frames, 246 columns (*)
Waveform	5 to 6 frames, about 4600 to 5500 columns	0 to 3 frames, at most 9 columns (*)

(*) What remains is legitimately empty: zoomed out past where this run's rolling history begins (nothing was ever decoded there), the track's fade-out at its end, and single columns at the playhead.

Strips are not attached; they show only the fixture library.

Review round 1

• A track change during a pending zoom no longer keeps the previous track. A reset counts as a zoom restart only if it is for the same track, the clock has reported no seek or track change since the zoom was requested, and in rolling mode it starts at the playhead. Any seek or track change the clock reports drops both layers and the reserve still being drawn. Layers now carry their track and validity in rolling mode too. Test: spectrogramTrackChangeDuringAZoomShowsNothingOfTheTrackBefore (the clock arriving first, and the data arriving first).
• Sparse waveform envelopes cover their own frames. Each point now covers the pixels whose centres fall inside its frames, clipped to the slice and to the detail. Points narrower than a pixel are placed as before. Test: waveformEditorSparseDetailCoversEachPointsFrames (alternating loud and silent buckets, slices that start and end inside buckets).
• Memory: the reserve is limited to the plot's logical width, so at a device pixel ratio of 2 it takes half the device-sized image (its rows stay at device height, which its frequency mapping needs). It is scaled up when drawn. The backdrop stays at full detail.

Review round 2

• Zoom restarts are recognised by identity, with no distance thresholds. A reset counts as the zoom's restart when:

  • its hop is that of the zoom being requested, or of a zoom the debounce sent to the backend since the zoom began;
  • it is for the track the zoom was requested on;
  • the clock's epoch hasn't changed since then.

  A UI seek calls holdAt(), which always bumps the epoch, so any seek counts as a break, however small. The 1.5 s and 0.75 s checks are gone, and the clock no longer drops the layers.

• Gapless and Repeat-One continuations keep the layers. Rolling layers follow the main canvas's rule: a data chunk with a new token and no reset continues the history. Only a reset that isn't the zoom's restart drops them. A continuation also carries a pending zoom's track (data side) and clock epoch (continueGaplessLocked) over to the successor. Layer times are counted from a stream origin, which absorbs m_gaplessPositionOffset whenever a reset clears it, so the layers stay correctly placed after a later zoom restart. Centered layers still check the track.

• Tests:

  • spectrogramLateZoomDataKeepsTheRollingHistory: an advancing clock, with the zoom's data 2.5 s late.
  • spectrogramSmallSeekDuringAZoomBreaksTheHistory: a 0.2 s seek while a zoom is pending.
  • spectrogramGaplessContinuationKeepsTheRollingHistory: gapless and Repeat-One, with the clock first and the data first. History is kept, and placement is checked after a zoom restart of the successor.

  spectrogramTrackChangeDuringAZoomShowsNothingOfTheTrackBefore still passes. All the new tests fail against round 1, and the placement check also fails without the stream origin.