Make popup windows fully visible on focus and attention-demand

[develop7]

Closes #1174

UPD: I've tested the fix and it works for me just fine, am using it daily.

Transient/popup windows are now repositioned fully on-screen when they appear, instead of being left wherever mutter clamped them — usually half-off the monitor edge, partially visible.

Popups (transients and dialogs) are real MetaWindows that mutter positions itself. PaperWM never scrolled them — they're rejected by add_filter so they're not in the clone container, meaning ensureViewport/move_to (which move cloneContainer) physically can't reach them. Two entry points now reposition them with move_frame:

Trigger	Before	After
Popup gains focus	focus_handler early-returned; popup stayed clamped	repositioned fully into the workArea
Popup demands attention but focus is denied (mutter focus-stealing prevention)	nothing — PaperWM connected no window-demands-attention handler	repositioned without stealing focus (the user's current window keeps focus, by design)

The demands-attention path is the common one for background-spawned popups (most are skip_taskbar, so gnome-shell's default WindowAttentionHandler bails on them too — they previously disappeared entirely).

How it's structured

The pure pieces live in a new shell-free popuputil.js so they're unit-testable with standalone GJS (tiling.js imports gi://Meta + the live shell, so it can't be loaded in plain gjs):

• workAreaToBounds — reconciles Space.workArea() (monitor-relative) to the screen-absolute coords get_frame_rect/move_frame use. Folding the monitor origin in here is the load-bearing detail: without it the popup jumps to the wrong monitor on multi-head.
• computeClampedPosition — clamps a frame into bounds; an oversize window pins to the leading edge rather than going negative.

tiling.js holds the thin shell-facing adapters: isPopupClass (derived from the canonical add_filter, minus sticky/scratch), ensureVisibleInWorkArea (clamp + move_frame), and positionPopupOnDemand (the demands-attention handler, no focus steal).

Caveats

• The focus path handles actual transients only (isTransient); the broader isPopupClass (transients + non-NORMAL dialogs) is used only on the new demands-attention path. Widening the focus branch too regressed side-effects for standalone non-NORMAL windows (topbar/overlay/selection), caught in review.
• move_frame on Wayland may be re-clamped by mutter's transient constraints on a later pass — the pure positioning logic is unit-tested, but the live move_frame behaviour needs manual verification in a gnome-shell session (can't be automated without a running compositor).

tests/test-popup-visibility.js (17 checks, gjs -m) covers the coordinate reconciliation, the clamp arithmetic including the oversize edge case, and a multi-monitor end-to-end.

---

Generated by /do on opencode (model zai-coding-plan/glm-5.2).

[develop7]

Hickey/Lowy Analysis

Post-implement structural review (two parallel sub-agents). Both lenses converged on the same three headline findings.

#	Lens	Finding	Disposition
1	Hickey+Lowy	isPopupClass re-derived the four tiling-eligibility booleans (transient, window_type, on_all_workspaces, scratch) that add_filter already inspects	Fixed in this PR (88ee77a) — delegates to add_filter
2	Hickey+Lowy	positionTransientOnDemand was a no-this instance method, and its name said "Transient" while the body gates on the wider isPopupClass	Fixed in this PR (ff6b555) — hoisted to module-level positionPopupOnDemand
3	Hickey+Lowy	workAreaToBounds left the canonical inline reconciliation (centerWindow, cycleWindowWidthDirection) unconverted	Fixed in this PR for cycleWindowWidthDirection (168f675); ⚠️ No-op for centerWindow — it straddles two coordinate spaces (monitor-relative for move_to, screen-absolute for easeScratch), so forcing the single-space helper regresses clarity
4	Hickey	duplicate non-NORMAL-type test cases (MODAL_DIALOG and UTILITY collapse to the same isNormalType:false) overclaimed distinct coverage	Fixed in this PR (b0bdfb0)
5	Hickey	ensureVisibleInWorkArea docstring asserted a monitor-agreement precondition without enforcing it	Fixed in this PR (88ee77a doc note, then enforced as a get_monitor() guard in 4d98346)
6	Hickey	"route the focus branch through positionPopupOnDemand to unify the gate"	⚠️ No-op — would add a redundant isPopupClass check; the focus branch already does a single discriminating check + direct ensureVisibleInWorkArea

Notes

• The pure/impure split (popuputil.js shell-free ↔ tiling.js adapter) is a clean seam (Lowy) — it tracks a real axis (GJS testability), interface is a stable rect-in/rect-out contract. Not a hidden re-completing concern.
• Workspace relocation of cross-workspace popups is correctly left to the existing insertWindow redirect — not re-complected here.
• The focus-vs-no-focus policy (Case A keeps focus / clears selection; Case B doesn't steal focus) is intentionally expressed as caller choice.

Subsequent /code-police pass refined one item further: it reverted the focus branch's isTransient→isPopupClass widening (cfd21ae) because non-transient non-NORMAL windows lost tiled-focus side effects (topbar/overlay/selection) — a regression hickey-lowy's predicate-consistency lens didn't catch. The broader isPopupClass now lives only on the new demands-attention path.