feat(sparql-anything): add SparqlAnythingConverter for chunked non-RDF to RDF conversion

[ddeboer]

Adds @lde/sparql-anything, the first piece of porting geonames-rdf (shell scripts driving the SPARQL Anything CLI) to LDE.

SparqlAnythingConverter runs the SPARQL Anything jar once per input chunk to bound memory use, then streams the per-chunk N-Triples outputs into a single file. Processes are spawned through an @lde/task-runner, so the same converter works on the host, in Docker, or anywhere else a TaskRunner is implemented.

Domain-agnostic API

The originally reviewed version took a required adminCodesFile – a GeoNames concept in the constructor of a package that exists to convert anything, and one that could not express the invocations passing no --load at all. It is now an optional, generic load. Folded in before the first publish, where the rename is free. The general rule this prompted is #783.

It stays a single string rather than the string[] the review suggested: SPARQL Anything's -l takes one path, and a file (default graph) is not interchangeable with a directory (one named graph per file), so an array would have been both unsupported and semantically different.

Safeguards ported from map.sh

• A non-zero exit aborts the whole conversion, so a crashed chunk cannot be silently dropped from the output.
• An empty or missing chunk output aborts it too. SPARQL Anything exits successfully when it cannot read or parse an input: it logs the problem, writes nothing and stops. map.sh guards this explicitly ([ -s … ]) after a run once shipped a geonames.nt missing every ontology triple while staying green.
• An empty chunk list is an error rather than an empty output: a chunking step that produced nothing has already failed.
• convert() resolves only once the concatenated output is flushed and closed. pipeline(…, { end: false }) resolves when the source ends, so an un-awaited output.end() would let convert() return with writes still pending.

Paths, and where per-run files live

A review of the above found that the file handling, not the conversion, was where the remaining defects were:

• Every interpolated path is quoted (shellQuote). Both runners execute through a shell – NativeTaskRunner spawns with shell: true, DockerTaskRunner runs sh -c – so a chunk name containing a space broke the command, and one containing ; or a backtick executed whatever followed.
• A required workDir – the runner's cwd, or its Docker mountDir. The generated query file used to be written to the host temp directory, which a container cannot see, so Docker never actually worked despite the documented claim. Query files and chunk outputs now go into a fresh subdirectory of workDir and are named relative to it, so one command works in both places.
• That subdirectory is removed when the run ends. Chunk outputs used to land beside the inputs and stay there, which also meant a later run whose process wrote nothing at all could pass the non-empty check on the previous run's triples.
• Inspecting a chunk output rethrows anything but ENOENT, so a permission problem is not reported as an empty conversion.
• A newline is written between concatenated chunks, so two triples cannot share a line if a chunk output does not end in one.

workDir is required rather than defaulted: only the caller knows what their runner can see, and a silent host-temp default is exactly what made the Docker failure so confusing.

shellQuote moves into @lde/task-runner

sparql-qlever already carried a private shellQuote, added after an apostrophe in 's-Hertogenbosch silently produced 0-triple imports. Rather than copy it, it moves to the layer both consumers share, with tests that round-trip awkward filenames through a real shell; sparql-qlever now imports it. Its coverage thresholds drop by 0.02pp because the helper it no longer holds was fully covered.

Also in this update

• Reference documentation moved to docs/reference/sparql-anything.md and registered in the sidebar and packages overview; the README is back to the npm-page shape. The docs site became the source of truth on main after this branch was opened.
• Manifest starts at 0.0.0, per the versioning guidance now on main – which supersedes this branch's own docs: start from 0.1.0 for new packages commit, dropped in the rebase.
• @lde/task-runner is depended on as ^0.2.13, matching the workspace version as @nx/dependency-checks requires; nx release rewrites it when task-runner is bumped.

Remaining review findings – concurrency, JVM heap and a CLI passthrough – stay in #782 and follow after this merge.

Note for the maintainer: @lde/sparql-anything is not on npm yet, so the release run on merge will fail to publish it until it is bootstrapped manually. Bootstrap after that release run, so the manifest already carries the released @lde/task-runner range.