Stability & correctness follow-ups from long-term macOS testing — fix registry

[AnyCPU]

Since the macOS glitches report (#101, now closed — thanks for the fast turnaround on that batch), I've kept using and stress-testing gogpu/ui in a real application, mostly on Intel + Retina macOS. That surfaced another round of issues — memory leaks, rendering staleness, and event-routing bugs — spanning both this repo and its dependencies (gogpu, gg, signals).

I've root-caused them, and for most of them I have working, tested fixes that have been running in my tree for a while. Rather than one giant report, I'm filing them as focused issues — each with reproduction, root cause, and a patch attached — a few at a time so it stays reviewable. This issue is the index; I'll keep the list below current as reports go up and as they get resolved.

Themes

• Widget lifecycle leaks — overlays and virtualized containers (ListView/GridView) skip mount/unmount steps, leaking signal subscriptions, goroutines, and cached GPU scenes.
• Rendering staleness — stale textures and damage-tracking edge cases: ghost rows on fast scroll, stale styling after selection, boundary-texture churn on resize.
• HiDPI / multi-display — device scale not re-synced when a window moves between displays of different scale.
• macOS event routing — phantom pointer events with screen coordinates misread as window coordinates; window-title mis-rendering on Intel Macs due to an ABI-gated AppKit enum.
• Threading contract — a documented goroutine-safe API whose call path writes UI-thread state without synchronization (with a trace this time, not just a symptom).

Already shipped

One report from this round has already landed upstream: gogpu/gg#418 (composite-glyph fan-out/cycle hardening + a variable-font guard), shipped in gg v0.50.4 — thanks for the quick adoption.

Reports

(checked = resolved upstream; the list grows as reports are filed)

gogpu/gogpu

• darwin: ignore pointer events not bound to the handling NSWindow (phantom hover/click/scroll from the desktop) gogpu#364
• darwin: NSTextAlignment constants use the arm64 ABI ordering unconditionally — window title renders flush right on Intel Macs gogpu#365
• browser: wheel handler leaves ScrollEvent X/Y at (0,0) — populate from the DOM event's offsetX/offsetY gogpu#366
• windows: wheel ScrollEvent X/Y are physical pixels on HiDPI — missing the DPI divide every other pointer event has gogpu#367
• darwin: Config.TabbingMode/TabbingIdentifier are silently no-ops (selectors declared but never registered) gogpu#383
• darwin: window title rendered twice (overlapping) when "Show Tab Bar" is enabled gogpu#384

gogpu/ui

• Overlays bypass MountTree/UnmountTree: signal bindings never form on push, and never release on dismiss (goroutine + heap leak) #171
• desktop: device scale never re-synced after a cross-display drag — window keeps rendering at the old display's scale (blurry / oversized) #172
• ListView: dropped row decorators never release their cached GPU scene — peak heap grows with scroll distance #173
• ListView row content is never mounted — signal bindings created in a row's Mount() silently never register #174
• No Window.Close teardown: animation pumper goroutine survives window close and keeps calling RequestRedraw; animPumper.Stop can silently miss its goroutine #175
• Window resize releases ALL boundary textures and force-dirties every widget — full-tree GPU texture recreation burst on every resize tick #176
• Stale ghost rows after ListView scroll: unbounded consecutive damage-aware frames cycle full-blit coverage out of the damage ring #177
• ListView static-selection click updates the row background but not baked "selected" content styling until an unrelated repaint #178
• app: pointer/wheel events reach widgets while the cursor is outside the window bounds #179
• examples/modular-compositor: module goroutines have no shutdown path (blocked forever at window close) #180
• GridView cell content is never mounted — Mount-time signal bindings are silently dead; proposed fix included #181
• linechart's "safe from any goroutine" contract writes render-thread state unguarded — proposal: route the invalidation through the scheduler #182
• Damage rects: logical vs physical pixel mismatches across the damage/present path — 5 findings (2 production-affecting, 3 in gogpu/gg) #195

coregx/signals

• signals: drop per-Subscribe goroutine (context.AfterFunc); Once-guard Unsubscribe coregx/signals#6

gogpu/gg

• Composite glyph resolution has no fan-out budget or cycle detection (DoS); gvar deltas corrupt composite points in variable fonts gg#418

gogpu/wgpu

• metal: multisample textures allocated with Shared storage on integrated Intel GPUs — MTLTextureDescriptor validation abort wgpu#271

Notes

• Some things I investigated turned out to be already fixed on main (e.g. the collapsible ghost-content issue from Gallary example: closing the Data Display Section does not cully clear the content #147 — confirmed fixed by the boundary-culling change) — those aren't re-filed.
• One suspected defect from the same testing round (ListView rows rendering ~400px above their
  slot, seen on v0.1.27) was re-checked against current main before filing: a headless
  reproduction harness (offscreen renderer, screen-origin assertions across 7 configurations
  including Retina scale and panel collapse) measures 0px stamping error — it appears fixed by
  the render-pipeline work since then, so no issue is filed for it.
• Everything below the fold in each report is self-contained: environment, repro, root cause with file/line citations against the current release, and a git apply-able patch. Happy to adjust format if something else is easier to work with.