import-nanopub-chain: follow the FORRT backbone, not just refersToNanopub

[annefou]

Problem

The constellation importer (scripts/import-nanopub-chain.py) discovered neighbours only through the curated KnowledgePixels npa:refersToNanopub graph. For real FORRT chains that graph materialises essentially one edge — CiTO ↔ Outcome — so a BFS entering from the CiTO stopped after 2 nodes and never reached the Study, Claim, AIDA or Quote.

Those steps are linked by domain predicates the network graph doesn't index:

Outcome --isOutcomeOf-->     Study
Study   --targetsClaim-->    Claim
Claim   --asAidaStatement--> <purl.org/aida/…>  --(asserted by)-->  AIDA
AIDA    --related-->         Quote

Fix

• backbone_neighbours() — reads every nanopub a node points at straight from its TriG, plus resolves the Claim→AIDA hop via a new aida-sentence-nanopub.rq query, then keeps only targets that are themselves FORRT chain steps (chain_step_kind()). Value-lists, templates, papers and other noise are dropped and never crawled, so this stays robust without hard-coding a predicate list.
• walk() merges these with the existing refersToNanopub neighbours (new edge relation backbone).

Verified

Against a published 6-step chain (white-shark geolocation), entering from the CiTO:

Before: Imported 2 nanopubs (CiTO + Outcome).
After: Imported 6 nanopubs, 8 edges — Quote, AIDA, Claim, Study, Outcome, CiTO, each classified by template.

No new dependencies (rdflib already required). This also flows downstream to the per-replication repos and the GRID4EARTH benchmark template on their next sync.

[annefou]

Closing — wrong layer and wrong approach.

The canonical constellation is science-live-platform (api/src/np/constellation.ts, served at /np/constellation), and it already walks the full FORRT backbone: it mines every nanopub URI from each TriG (catching isOutcomeOf / targetsClaim) and bridges the Claim→AIDA gap via discoverAidaStatementNeighbours + the existing aida-statement-nanopub.rq query (committed 2026-05-31). Verified on dev: the white-shark CiTO returns all 6 steps (cito, outcome, study, claim, aida, quote).

This PR instead added hand-coded link-walking + heuristic step-type classification to the legacy import-nanopub-chain.py, which is exactly what that script's docstring says discovery should NOT do ("driven by the platform's pre-built SPARQL queries, rather than by hand-coded link-walking + heuristic step-type classification"), and it used the wrong placeholder convention (${aidaUri} instead of ?_name). If the legacy fallback ever needs to reach AIDA/Quote, the right fix is a query-driven re-sync that re-copies the platform's aida-statement-nanopub.rq and mirrors constellation.ts's TriG URI-mining + AIDA bridge — not these heuristics.