Drag-view alpha flattened to white: koncept drag glyph shows white corner notches on JavaFX 27-ea

[knowledge-graphlet]

Symptom

Every drag image built by KonceptDragGlyph shows opaque white notches at the four corners of the rounded pill — visible when dragging over a non-white backdrop (the navigator panels), invisible over white (the assistant compose area), which is why the glyph looked clean where it was first exercised.

Analysis — the transparency is correct at the source

The pill's corners are genuinely transparent in the produced image:

• glyphStyle() paints only the pill fill and border (-fx-background-radius: 6), no white layer;
• render() snapshots with SnapshotParameters.setFill(Color.TRANSPARENT) into a fresh WritableImage (transparent by construction).

So the alpha is being flattened to white downstream of Dragboard.setDragView(Image, …) — in the FX-image-to-native-drag-image conversion. Komet currently runs the pinned JavaFX 27-ea (see komet-bom's JavaFX 27-ea+24 pin rationale); transparent drag views have historically worked on JavaFX/macOS, so an ea-line regression in the drag-view conversion is the prime suspect. The legacy DragImageMaker snapshot path would flatten identically, but its images were full rectangles, so nothing was visible.

Suggested handling

• Re-test on JavaFX 27 GA (expected 2026-09-15, when the ea pin is scheduled to be dropped). If GA restores alpha, close.
• If GA still flattens: reduce exposure by minimizing transparent margin in the glyph (the notches are only the 6px corner radius), or accept the artifact and reference the upstream JBS issue once identified.
• Not proposed: pre-compositing the corners onto an assumed backdrop — the backdrop is unknowable and both light and dark panels are in play.

Acceptance

Dragging any koncept over the navigator's grey backdrop shows the rounded pill with no white corner notches.

[knowledge-graphlet]

Checked the newest available ea build against the pin: openjdk/jfx 27+24...27+25 contains seven commits — 8268657 (looked-up color fix) plus its backout 8388140, a GridPane sizing fix, RichTextArea attribute additions (8366198), two NPE fixes, and a GHA change. None touch macOS glass, drag-and-drop, drag image, or clipboard conversion, so bumping the pin to 27-ea+25 cannot affect this artifact.

This also weakens the 27-ea-regression hypothesis: no drag-layer churn in the ea line, and no matching JBS issue found (the known white-square drag-image bugs are all embedded contexts — JFXPanel/SWT — or JavaFX 2.x era). The stronger hypothesis is now that macOS glass has flattened drag-view alpha all along, unseen over white surfaces.

Discriminator: drag over a grey backdrop in a komet-desktop installer built on the 26.x line (pre-pin checkpoint releases exist). Corners there too → perennial platform behavior → minimal upstream repro + JBS report is the durable path; clean there → ea regression after all, report upstream before the 27 RC window closes (GA 2026-09-15).

[knowledge-graphlet]

Interim landed: the drag pill renders square (radius 0) so the image has no transparent pixels — no notches over any backdrop. A pixel test first PROVED the rounded pill's corners were alpha-0 in the produced image (the flattening is downstream), then was flipped to pin the interim and documents the revert: restore radius 6 and the transparent-corner assertion when this issue resolves.

[knowledge-graphlet]

Frame-refinement update (2026-07-24, komet 9298b6f8/3d29949d/227f1318): bevel tones now derive from ONE base (#c8d6e6, the shared floating grey — light = interpolate(WHITE, 0.85), shade = deriveColor sat×1.6 brightness×0.82); lit top/left edge trimmed to 1px against the 2px shadow (KEC verdict: improved, stands). The retina-fuzz dial is RESOLVED: drag entry points detect Screen.getOutputScaleX() and deliver the full-density raster via the JavaFX name@2x.png resolution-variant loader — validated by pixel forensics on a full-resolution capture of a live drag (35-logical-px card at 281×69 device px, ratio 1.97, single-device-pixel edge transitions; an OS-upscaled 1× image would show interpolated rows). The 1× supersampled path remains the fallback. Still open here: the alpha-flattening discriminator (drag on a pre-pin 26.x checkpoint installer) — with an opaque, density-correct frame that now looks intentional, the urgency is lower, but transparency restoration at revert still hinges on it.