Every component is now two modules sharing a basename:
- src/elements/leaflet-foo.ts -- `export default class LeafletFooElement
extends WithProps(...)`, the class only, no customElements.define, no
side effects. src/elements/index.ts is an order-free barrel.
- src/components/leaflet-foo.ts -- three lines: import the class, define
the tag, re-export. Importing this (or src/index.ts) registers the tag.
Plugin authors can now import a class without triggering the built-in
define, to subclass it or register it under a different tag name. Package
subpath exports `leaflet-components/elements`,
`leaflet-components/elements/leaflet-foo.js`, and
`leaflet-components/components/leaflet-foo.js` map onto dist/; jsr.json
gains an `./elements` entry.
The load-bearing ordering moves from the export statements in src/index.ts
to its `./components/*` side-effect imports (the elements barrel carries no
define, so its order is free). Classes are renamed LeafletFoo ->
LeafletFooElement, including in the HTMLElementTagNameMap augmentation.
Docs (CLAUDE.md, docs/, README.md) updated for the new layout.
Two parent-looks-for-child patterns found and removed, per the
child-emits-events-parent-never-reaches-in principle:
- leaflet-control-layers: #childLayers() queried children directly via
querySelectorAll + reading their leafletObject/attributes, in two call
sites. Verified empirically that a parent's connectedCallback always
completes before a freshly-connected child's does (even for an
already-built subtree attached in one shot), which means both call
sites always ran against unpopulated children -- dead code. Deleted;
#onChildRegister (already event-driven) was doing all the real work.
- leaflet-polygon/leaflet-polyline: #coords() queried <leaflet-line>
children via querySelectorAll + read their .latlng property directly,
using a MutationObserver + a payload-less event only as a "something
changed, rescan everyone" signal. Rewritten to be purely event-driven:
<leaflet-line> now fires leaflet-line-sync (connect + every lat/lng
change, carrying its own position) and leaflet-line-remove (disconnect)
on itself; a new shared VertexTracker (src/core/vertex-tracker.ts)
turns that event stream into an ordered coordinate list, using
compareDocumentPosition only to place a newly-registered vertex at its
real document position rather than assuming registration order matches
DOM order. No querySelectorAll, no MutationObserver, no reading a
child's property.
Testing the polygon rewrite in an actual browser (not jsdom) surfaced two
real bugs, both fixed:
1. disconnectedCallback fires *after* a node is already detached from its
parent, so leaflet-line's removal event had nowhere to bubble to.
Fixed by caching parentNode while still connected and dispatching from
that cached reference instead of from the (by-then-detached) node.
2. A load-order hazard affecting every "parent listens for a specific
child tag's announcement" relationship in this codebase:
customElements.define() upgrades every matching element already
parsed into the page immediately and synchronously, so whichever tag
gets defined first wins a race -- a child tag defined before its
listening parent fires its one-shot connect-time announcement into a
parent that doesn't exist yet, and it's lost for good. This broke
polygon/polyline (introduced by this rewrite, since leaflet-line was
exported before leaflet-polygon) and, independently, was already
silently broken for leaflet-control-layers (a base/overlay layer
leaked directly onto the map, bypassing the control entirely) and
latently for leaflet-layer-group/leaflet-feature-group. src/index.ts's
export order now encodes and documents the real dependency hierarchy
(map -> containers that listen for child announcements -> concrete
layer types -> their own children).
Added test/load-order.test.ts, which is structurally different from
every other test file: it never statically imports a component
module, building the DOM with plain undefined elements first and only
dynamically import()ing src/index.ts afterward -- reproducing a real
page's actual load order, which every other test's "define everything
first" pattern cannot catch. Confirmed it fails against the old
ordering and passes against the fix.
Buddy