zelph 0.9.7

🧠 Neural Networks in the Graph

zelph can now embed neural networks directly inside its semantic network — neurons are ordinary graph nodes, synapses are weighted edges (invisible to the reasoning engine), and layers are ordinary sets. Sub-graphs compile into feed-forward networks on demand, are trained via SGD, and write their weights back into the graph.

• New ≈ rule condition operator: rules can consult a trained network — guard mode verifies existing facts against the model, generator mode proposes new bindings above a confidence threshold. The network's confidence becomes the deduced fact's probability, stored in the same weight store as ordinary fact probabilities.
• 12 new Janet functions (zelph/nn-connect, zelph/nn-compile, zelph/nn-train-nodes, zelph/nn-eval-nodes, zelph/approx, …).
• New standard library script nn.zph with high-level helpers (nn/link-predictor goes from graph facts to a rule-usable network in one call).
• Proof of concept on real Wikidata data: examples/neural/nn-wikidata-demo.zph learns country→continent (P30) links, verifies existing facts, and proposes candidates — with confidences flowing back into the graph.
• New documentation page: Neural Networks in the Graph.

📚 Standard Library and .import Resolution

• zelph now ships a standard library (arithmetic, binary-arithmetic, nn, sparql, plus an examples/ tree), installed system-wide by all packages (e.g. /usr/share/zelph on Linux) and included as stdlib/ in the release archives.
• .import resolves scripts against the working directory first, then the standard library; the .zph extension is optional: .import sparql, .import examples/english. Search paths can be overridden via $ZELPH_STDLIB.
• Scripts passed on the command line resolve the same way (zelph examples/english), and the packaged test suites are now hermetic (Chocolatey and Homebrew run them again at install/test time).

🔢 Number Literals (&)

• New & prefix for number literals: &42 is always decimal input, regardless of the internal representation. It delegates to the redefinable Janet function zelph/number — the numeric representation lives entirely in scripts, not in the engine.
• New binary-arithmetic.zph: multi-digit addition in base 2 from the 16 hand-written full-adder facts — no generated lookup table, pure native zelph syntax.

🗂 Node Clusters (Transactional Workspaces)

• New commands .cluster, .cluster-drop, .cluster-merge: record every node created during an experiment and roll it back (or commit it) with a single command. Pre-existing knowledge is never touched.

⚡ Reasoning Performance

• Subject/object-anchored candidate snapshots now also apply to conditions with concrete constants. Previously such conditions scanned the entire relation extent — e.g. all ~15 million P31 facts of a Wikidata dump per rule application. A representative Wikidata rule dropped from 29 s to 1.5 s.
• The template-node detection that protects this fast path is now structurally precise (deep variable check), keeping cons-list/arithmetic rules correct.

🐛 Fixes

• Variable display names are no longer "stolen" when a later rule reuses the same variable names; .list-rules no longer degrades earlier rules to «??» monstrosities.
• Unnamed variable nodes are never expanded structurally in output anymore.
• REPL no longer gets stuck after a Janet error inside a % block.

All 108 test cases pass on Linux, macOS, and Windows. Debian packages now additionally install the standard library to /usr/share/zelph.