Skip to content

Release/1.8.0 - #50

Merged
ThHanke merged 45 commits into
mainfrom
release/1.8.0
Nov 11, 2026
Merged

ThHanke merged 45 commits into
mainfrom
release/1.8.0

Conversation

@ThHanke

@ThHanke ThHanke commented Nov 11, 2026

Copy link
Copy Markdown
Owner

No description provided.

Thomas Hanke added 30 commits November 10, 2026 21:46
…in shapes only as fallback

MCP tools used to answer while ontology autoload, ?rdfUrl= and startup shapes were
still loading, so validateGraph and runReasoning saw a half-loaded graph and runs
differed between repetitions. The MCP entry point now rejects every call until
startup is complete (120 s safety valve), and window.__startupComplete() exposes
the state for e2e tests.

The built-in ontology-quality shapes load only when neither ?shaclShapes= nor the
configured URL names shapes. The factory default of the shapes URL is empty.

An empty shapes URL used to mean "disable auto-loading"; it now means "use the
built-in shapes". Turn SHACL off with the SHACL toggle.
…logy on demand

rdf, rdfs, owl and prov are always fetched from w3.org first; the copies in
public/w3-vocabularies are used only when that fails (blocked, offline).

The PyodideSemanticWorkflow ontology is reasoning input, so startup no longer
loads it. The catalog still loads at startup; the ontology joins urn:vg:ontologies
when a workflow is dropped on the canvas. ?workflowCatalog=off skips the startup
catalog load for headless runs.
… the reasoner API; one explain per graph state

Reasoner
- Pin rdf-reasoner-konclude 0.9.1 (fixes blank-node ids returned by the reasoner, adds the
  repair() justification seed). The WASM and worker in public/ come from the package.
- A stalled or crashed reasoner is a failed run: SHACL is skipped, the result status is
  error and the MCP tool runReasoning returns success:false.
- Violations without a focus node carry their shape as nodeId.
- Dataset and Turtle export no longer drop literals that differ only by language or
  datatype (the de-duplication key ignored both; PMDco 3b98aad lost 7 triples).

Repair, weakening
- src/mcp/tools/repairAdapter.ts maps repair(), weaken() and assessEdit() onto the existing
  RepairSuggestion shape; a worker command suggestRepairs serves explainDiagnostics and the
  repair panel. Set size dominates the cost policy, then assertion over schema axiom.
- Deleted what the reasoner now owns: axiomWeakening.ts, the greedy hitting set and
  minimality code, the verifyRepair command and verifyRepairMatch.ts. The SHACL repair
  pipeline and the shared types stay. Removal by exact quad identity is covered by
  removalIdentity.integration.test.ts and the library's assessEdit tests.

explainDiagnostics
- No laconic explanations (they only fed brief wording; weaken() splits conjuncts itself).
- Justifications come from the repair search; one whole-axiom explain per graph state,
  cached per worker runtime on reasoningInputGeneration and handed to repair() as a seed.
- unsatisfiableClasses is not computed for an inconsistent ontology (every class is
  unsatisfiable there); unsatisfiableClassesSkipped says so and the UI shows a toast.
- includeTimings adds timingsMs (stage and reasoner-call durations).
- PMDco plus one clash in headless Chromium: 513 s before, 128 s now.
CHANGELOG: laconic removal and why, explainDiagnostics changes and speedup, unsatisfiable
classes skipped for inconsistent ontologies, export de-duplication fix, stall handling.
README: repair and weakening on the reasoner API, startup and shape-loading behaviour,
?workflowCatalog=off.
…pace map

The tool took only full IRIs and answered "NOT entailed" (success: true) for anything
else, so a prefixed name copied from a queryGraph row looked like an unentailed fact.
Prefixed names are now expanded with the shared namespace map, full IRIs pass unchanged,
and an unknown prefix or a bare name is an error, as in the other tools.
… at scale

- user-story-pmdco-chemical-composition: the ISWC demo scenario with assertions. Data, PMD Core
  and the auto-shapes load through URL parameters, reasoning finds the one some_silicon
  violation, the BandGap fix makes the data conform. Pinned to the June 2026 PMD Core versions,
  where the violation exists.
- reasoning-inconsistency: explainDiagnostics -> suggested repair -> removeLink -> consistent,
  plus explainEntailment with full IRIs, prefixed names and the two error cases.
- pmdco-repair-scale (manual, PMDCO_SCALE=1): applies the suggested repair and reasons again,
  checks the ontology is loaded when startup completes, PMDCO_URL selects the ontology file.
…de timing test

The tool description and manifest say deletion repairs have kind "delete", but the adapter
left kind unset (the old code did too). Weakenings keep kind "weaken"; no consumer relies on
a missing kind (all test for "weaken" only).

The node timing test (pmdcoScale.manual.test.ts) is removed: explainDiagnostics({ includeTimings })
and the manual e2e spec e2e/pmdco-repair-scale.spec.ts cover the same measurements.
…d for runtime

- owl:oneOf realization: skipped since v0.3.0 ("does not fire"); passes on 0.9.1.
- P1-4 import and TriG round trip: run on the bundled bench-pmdco.ttl by default (P14_FILE
  still selects another file). They fail with 4 triples missing without the export fix.
- PMDco repair at scale (e2e): no longer behind PMDCO_SCALE. About 2.5 minutes, stable across
  runs (explainDiagnostics 126-128 s in headless Chromium).

Also replace a literal NUL byte in computeRepairs.ts (the axiom key separator) with '\u0000',
which made the file binary to grep.
…er as local names

A local name (the part after the last '#' or '/') is not unique across namespaces and cannot be
passed back to a tool. Messages, rationales, the repair brief and summaries now use a prefix from
the namespace map, or the full IRI when none is registered (src/utils/iriAbbreviation.ts). The
worker uses its own namespace map; main-thread callers pass the registry. The canvas label
substitution matches the new text and no longer falls back to a local name.
… validation scope

- Performance line: ranges over 5 fresh runs on 0.9.1; the PMDCO figure is corrected (2.6 s was not
  reproducible); LUBM-1 reported as a range.
- "What is validated": data graph plus inferred graph, loaded ontologies are not validated, and what
  that means for sh:closed shapes; SPARQL UPDATE scoping note; notes for agents under the tool table
  (startup rejection, IRI rules, explainDiagnostics output); tool count 43 -> 42.
- Startup order includes the shapes rule and the rejection note.
- Benchmark and demo drivers wait for window.__startupComplete() before their first tool call;
  bench-paper-table2.mjs gains --only and --fresh-browser.
Version 1.8.0 in package.json, the lockfile and CITATION.cff (which still said 1.5.0).
CHANGELOG: dated 1.8.0 section with upgrade notes (inferred counts change with the reasoner),
breaking changes (MCP calls rejected until startup completes, empty shapes URL, laconic removal,
justification and text output changes), reasoning, startup and loading, bug fixes, tests and tooling.
27 vulnerabilities (1 critical, 13 high, 13 moderate) down to 7 moderate. Lockfile only; package.json
and the pins (rdf-reasoner-konclude 0.9.1, n3 1.26.0) are unchanged. The remaining 7 need
`npm audit fix --force` (breaking major upgrades) and are left alone.
…ine, export dedup key, contract changes, verification pitfalls, shell and git slips)
… reasoner chain

Pin 0.9.3 (reads urn:vg:bnode:* IRIs itself, constructor excludeGraphs).
Add spike tests showing the real store works without the reasoningBase
copy. Ignore graphify-out/.
…; LUBM bench with a valid RDF TBox

reasoner-complete carries copy, consistency, materialize and write-back
durations; complete carries the post-reasoning emission steps. The new
bench-reasoner-phases script waits for the main thread to go idle before
timing. bench-lubm-tbox.nt (RDF) replaces the OWL/XML TBox the RDF/XML
parser could not read.
…k individual, pizza, PMDco, explain/repair outputs)

Real skolemized store with constructor excludeGraphs gives the same
inferences as the reasoningBase copy; outputs carry no blank nodes or
unknown skolem IRIs; no fnv1a32 collision among PMDco blank labels.
skolemReasoning: the stored form now reasons correctly on 0.9.2+.
… or inferred write-back

DlReasoner passes the live store to rdf-reasoner-konclude (0.9.3 reads
urn:vg:bnode:* IRIs as blank nodes and returns them in stored form) with
excludeGraphs set once at construction. The package writes the inferred
graph itself; the app keeps the delta for emission and SHACL only.
A stale run drops the inferred graph the package just wrote and reports
the previous run's inferred quads as removed. Graph counts for the
reasoner-managed graphs are recounted after every reasoner command.
reasoningBase, deskolemize and storedValue are removed; the spike test
for the rejected native-blank-node design is deleted.
… RSS per process type

The dev server pre-bundles rdf-reasoner-konclude and only re-checks on start; the script now refuses to run when the bundle is older than the installed package. Each run also reports VmHWM per Chromium process type.
…when the worker result is in

The data provider marks a triple as inferred from the quad's own graph, so the
emission handler no longer needs a separate fetch of the inferred graph.
handleApplyInferred is removed. The worker posts the inferred-subjects emission
after the reasoning result, so the result no longer queues behind the page's
long link and render work. LUBM-1 runReasoning: 126 s -> 12 s wall, worker 9 s.
Benchmark scripts: idle threshold (IDLE_MS), stage timeline, canvas-link scaling
and a main-thread timeline profile.
…ctodia createLinks

Patch-package: onLinkInfoLoaded builds an IRI -> elements map and createLinks looks sources and
targets up instead of scanning the whole diagram twice per link (links x elements).
…ta and links per chunk

A load creates elements in about 8 chunks (300 to 3000 each), requesting element data and links after
each and yielding a frame, so no single task covers every element; a set of queued IRIs keeps a second
emission from duplicating them. Many small chunks cost more in total than they save (each re-renders
all existing elements), hence the scaling. Longest task at 8000 individuals: 14 s -> 5 s, total unchanged.
bench-canvas-links: heartbeat-based settle detection.
…long-task stats and time to settle

LUBM-1 after reasoning: longest task 40 s -> 16 s; total catch-up is longer (about 83 s -> 108 s)
because every chunk requests links against all elements. bench-reasoner-phases resets the counters
before the call and reports time to settle from the call start.
…nput as RDF; phase benchmark takes a file list
Thomas Hanke added 15 commits November 11, 2026 16:53
…r edited elements

Reactodia mounted an authoring decorator, a portal with its own container, for every entity element
whenever the editor is in authoring mode; React registers its whole delegated listener set (about 138)
on each portal container. LUBM-1 had 17,174 of them: 2.37M listeners and about 50 DOM nodes each.
Patch: decorate the hovered element, selected elements and those with authoring, validation or
temporary state. Authoring mode and all actions are unchanged.
…hide the duplicate halo delete

Hover is tracked with one delegated mouseover on document (the pointerMove canvas event is a drag
event). The halo's delete button duplicated the entity's inline Delete and is hidden; the component
stays mounted for the Delete hotkey.
Collapsed card styles move from per-node inline style objects to .vg-card
rules; the namespace colour bar is a ::before and the content wrapper is
merged. LUBM-like bench (1000 individuals, 2000 links, headless RSS, no
layout): 31383 -> 30383 DOM nodes, slope 0.342 -> 0.323 MB/individual.
Unselected links on large canvases render only the path and hit area: no HTML
label node, no label measuring, no per-link link-type subscription. Selected
links keep the full template. Bench (1000 individuals, 2000 links, headless
RSS, no layout): 30383 -> 24893 DOM nodes, slope 0.323 -> 0.264 MB/individual.
… layout and URL memory profiling

Template, decoration or patch changes must be measured with bench-canvas-links.mjs before merge;
reference per-individual cost and the lazy-mount rule for per-element portals.
…RL memory scripts

The guard loads data through ?rdfUrl= (loadRdf stages entities as authoring
additions and mounts decorators) and stays under largeGraphThreshold so the
entities are not folded into groups. Budgets are set from measured values
(about 30 DOM nodes per individual with labelled links) plus 25%. CLAUDE.md
numbers corrected.
…job copies the PMDco bench input

importRoundTrip and noCopyReasoner read public/bench-pmdco.ttl and
public/bench-pizza.owl, which are git-ignored local files, so they failed in
CI. They now use src/__tests__/fixtures/pmdco.ttl and a new pizza.owl fixture.
The e2e job copies the PMDco fixture to public/ before the build for the
pmdco-repair-scale spec.
@ThHanke
ThHanke merged commit bf0a5f3 into main Nov 11, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant