feat(pan-zoom): configurable zoom-out for the double-click toggle

[webard]

What

dblClickZoomOutLevel — where a zoomOnDoubleClick: 'toggle' double-click zooms out to when it has no remembered viewport to restore.

flowCanvas({
  zoomOnDoubleClick: 'toggle',
  dblClickZoomLevel: 1,        // double-click → 100% about the cursor
  dblClickZoomOutLevel: 'fit', // double-click again → the whole graph
})

Why

The toggle's headline behaviour — second double-click restores the exact viewport the first one left — is the right answer whenever there is a viewport to restore. The branch with nothing remembered is reached more often than it sounds: the user got to this zoom by wheel or setViewport(), or panned after zooming in, which drops the memory by design (rememberedViewport = null on any user-sourced 'start').

That branch went to minZoom, hard-coded. The only lever on it was minZoom itself — the same floor the wheel and pinch answer to — so "zoom out to something sensible on double-click" and "let the wheel go down to 0.5" could not both be had.

Where this came from: a workflow editor at dblClickZoomLevel: 1. The gesture reads as closer, then show me all of it — and "all of it" is the graph's own extent, not an arbitrary floor that frames nothing in particular.

Shape

• 'min' (default) — minZoom about the cursor, byte-for-byte the old behaviour.
• 'fit' — the viewport fitView() computes for every visible node. Falls back to 'min' when there is nothing to fit (empty canvas, nothing measured yet), so the gesture never goes dead.
• number — a fixed level about the cursor, clamped to [minZoom, maxZoom].

A remembered viewport still wins over all three; the option only names the fallback.

pan-zoom knows nothing about nodes, so 'fit' asks for the viewport through a new getFitViewport?: () => Viewport | null option, wired in _initPanZoom from getNodesBounds() + getViewportForBounds() — the same pair fitView() uses, with the same DEFAULT_FIT_PADDING. It is passed to resolveDblClickZoom as a thunk: most double-clicks zoom in, and walking every node's bounds for a branch that will not run is work nobody asked for.

The no-headroom guard now measures against whatever the toggle zooms out to rather than always minZoom, so dblClickZoomLevel: 1.5 with dblClickZoomOutLevel: 1.5 keeps d3's stepped handler instead of installing a gesture that stalls on its second half — same rule as the existing level <= minZoom case. 'fit' is measured at gesture time and can't be checked statically, so it is held to the 'min' floor and the runtime fallback covers the rest.

Tests

src/core/pan-zoom.test.ts — eight cases on the pure resolveDblClickZoom: 'fit' returns the fitted viewport pan and all; the fallback to minZoom when there is nothing to fit; the thunk is not called on a zoom-in; a remembered viewport still wins; a numeric level zooms out about the cursor and clamps to minZoom; an out-level with no room below returns the current viewport by reference (the existing "honest no-op" contract); and the default matches the pre-option behaviour exactly.

src/core/pan-zoom-dblclick.test.ts — four in the wiring suite, two driving a real DOM double-click through the d3 transition: 'fit' settles on the supplied viewport, no measuring while the gesture zooms in, and the attach/skip decision for a numeric out-level.

npm run test — 187 files, 3044 tests, all passing. Five of the new cases fail against dev without the source change.

Not included

dist/ is not rebuilt — source, tests and docs only. No version bump, no CHANGELOG.md entry, no new dependencies.

Independent of #82; both branch off dev and touch different files.