Bound the tree children query and CONSTRUCT only what a node renders

[namedgraph]

The defect

ldh:tree-children-query emitted an unbounded DESCRIBE ?child, so opening the drawer tree fetched the full description of every child of a container. It fires on every page render, not on a click — ldh:NavigationUpdate walks down to the current document and fetches each level on the way.

Measured against a dataspace of ~158k documents, one entity per named graph:

container	children	DESCRIBE	CONSTRUCT
taxonomies/colors/	275	202 KB	—
sets/	28,356	20.5 MB	6.6 MB
parts/	64,647	502	18.1 MB

Both large containers exceeded the response limit, so the tree answered 502 on every page beneath them.

The fix

A node renders four things — @rdf:about, a label, rdf:type for the icon, and whether the tree relation is present for the disclosure — so a CONSTRUCT of just those is both correct and several times smaller. Making the size independent of container size needs the bound as well: ldh:tree-page-size (default 1000), which an app stylesheet can raise.

A bound alone would break the descent. ldh:doctree-descend picks its next step out of the rendered children and falls through silently when it is absent, which with a LIMIT it usually would be. The query now takes the document being opened to and includes the child leading to it whatever else the page holds.

Scope

The SKOS trees from the editor/taxonomy package call this same function and inherit the bound with no package change. Verified on a live instance: top concepts and narrower concepts both fetch through the new CONSTRUCT.

The package's own ancestor-path walk is deliberately left alone — it is a SELECT DISTINCT ?parent over a VALUES frontier, already minimal, and a LIMIT there would silently stop the tree opening to the target.

Tests

Three places in ui-tests matched the query by its verb and broke — two of them silently, which is why they are a separate commit:

• tree-children-failure routes the injected 403 on that pattern; a route matching nothing fails open, so the refusal never happened and every assertion passed against a tree that was never refused anything.
• concept-tree's counts.down stopped counting, so the DOM is the cache compared 0 against 0.
• lib/console allowed the 502 the unlimited DESCRIBE produced — now removed rather than updated, since an allowance that outlives its defect is a blind spot over exactly the request this changes.

Two new specs cover the fix, with the negative control that makes the second worth having:

run	container expanded	target active	children
baseline	yes	yes	1001
LIMIT 0	yes	yes	1
LIMIT 0, path branch removed	no row	no	0

The suite was not run end to end: it requires an instance where nothing is readable by default, and the stack available granted anonymous read across the dataspace, which its preflight rejects by design.

🤖 Generated with Claude Code