Zoom effects: elevate and animate the zoomed pane

[nedtwigg]

Makes pane zoom a real presentation layer instead of a geometry special case, and couples it to passthrough focus.

What changed

• Zoom is an animated, elevated layer. The zoomed leaf tweens into a wall-sized rect inset by half the pane-header height (15px), riding LATH_LAYER_ELEVATED — one of three layer bands (TILED / DYING / ELEVATED) the animator now threads through to the host. This replaces the old zoomedIdRef special case that skipped geometry for one leaf. The exposed perimeter plus a blurred app-background shadow make the stacking relationship visible.
• Zoom follows passthrough focus. Any focus transition off the zoom owner starts its return to the tiled layout; the zoomed pane's unzoom control is highlighted.
• New panes get focus after a manual split.
• Browser hover feedback. The agent-browser stream frame paints provisionally (CSS-resolution, immediate) while the crisp device-resolution host screenshot replaces it, so hover/animation feels live instead of stepping at capture rate.
• L-shaped desktop tutorial layout, seeded deterministically by the playground route.

Notes for review

The last commit is a /simplify pass over the branch (no behavior change) — dead ZoomedContext removal, one releaseZoomExcept helper in place of four copied guards, a single emitFrame path, and a dedup hash that stops copying ~100KB of base64 per frame.

Specs updated alongside: tiling-engine.md (layer bands, zoom geometry), layout.md (zoom/focus coupling), dor-browser.md (provisional paint), tutorial.md.

Test plan

pnpm test green: spec-lint (23 specs), lib 1063, website 51, standalone 47, plus dor/server.

Known follow-up, deliberately not in scope: during sustained pointer input the crisp screenshot loop starts captures whose results are then dropped as stale against newer provisional paints. Fixing it means gating take() on the provisional window plus an explicit trailing flush — a pacing change that needs its own verification, since a naive gate would strand the surface on the provisional frame when the stream goes quiet.

🤖 Generated with Claude Code

[cloudflare-workers-and-pages]

Deploying mouseterm with    Cloudflare Pages

Latest commit:	fa18c02
Status:	✅  Deploy successful!
Preview URL:	https://ff931b72.mouseterm.pages.dev
Branch Preview URL:	https://zoom-effects.mouseterm.pages.dev

View logs

[dormouse-bot]

One correctness finding on the two-stage browser paint; the zoom-as-layer rework and the focus↔zoom coupling both trace cleanly and the specs are accurate.

Provisional paint can strand the pane on the blurry frame at rest. The screenshot loop skips a decode when a capture's bytes match lastDrawnKey (agent-browser-screenshot-loop.ts display), but lastDrawnKey only ever updates on a crisp draw (line 83). drawProvisionalFrame → paintBitmap overwrites the canvas with the blurry CSS-resolution frame without touching lastDrawnKey, so the loop's dedup no longer reflects what's on screen.

Repro (HiDPI, host has agentBrowserScreenshot):

1. Crisp frame C is drawn — lastDrawnKey = C, canvas is sharp.
2. Hover a control → page changes → provisional blurry paint; a crisp capture starts.
3. Continued pointer jitter bumps provisionalPaintGeneration, so that in-flight crisp is dropped as stale (the mechanism you flagged in the follow-up note). lastDrawnKey stays C.
4. Un-hover → resting pixels equal C again → provisional paints blurry C; the coalesced crisp captures C.
5. That crisp is not stale, but display computes key(C) === lastDrawnKey and returns early — no draw. The stream is quiet, so the pane stays on the blurry provisional C indefinitely.

This is distinct from the acknowledged pacing follow-up, and would defeat its planned fix: even an explicit trailing crisp flush is swallowed by this byte-dedup when the resting frame matches the last crisp draw. The existing drawGeneration mechanism (bumped on re-attach to force a repaint of a blank canvas over identical bytes) is exactly the right tool here — a provisional paint is the same situation, so bumping it when a provisional lands makes the next crisp always repaint over the blur. Inline suggestion below.

Reviewed and clean

• ZoomedContext → ZoomedIdContext rename: all consumers migrated; no surviving import of the removed symbol anywhere in the tree.
• Animator layer bands (Math.max(from.layer, to.layer) during tween, to.layer when settled, framesClose now comparing layer, markDying using Math.max(cur.layer, LATH_LAYER_DYING)): consistent across rise / hold-during-lower / snap / reduced-motion / dying-while-elevated.
• Focus↔zoom coupling: releaseZoomExcept correctly threaded through selectPane / selectDoor / enterTerminalMode / exitTerminalMode; re-selecting the zoom owner no-ops.
• frameMetadata requiring both dims > 0, and the inline PANE_HEADER_STYLE replacing the removed .lath-leaf-header CSS: no consumer regressed.
• The desktop-playground L-shape seed matches LathPersistedLayout and is validated by the new test; specs (toDevice, PANE_HEADER_HEIGHT_PX, LATH_ZOOM_*, presentationTargets, LATH_LAYER_*) point at symbols that exist.

Minor, non-blocking: the large-frame dedup hashes the full JSON envelope (isDuplicateFrame(raw)) on the idle path but the base64 body + dims (isDuplicateFrame(msg.data, metadata)) on the wantFrameData path, so a byte-identical heartbeat straddling a wantFrameData toggle costs one redundant crisp capture. Benign and self-correcting — there's no clean fix without parsing on the idle path, so probably not worth changing.