MeTTa TS 1.2.0-experimental.0

MeTTa TS 1.2.0-experimental.0

This is an experimental prerelease on the experimental npm dist-tag, not the
default latest. It carries the in-progress Minimal MeTTa runtime and the
Grounded V2 operation protocol for early adopters; the surface may still change
before 1.2.0. Install it explicitly:

npm install @metta-ts/core@experimental
npm install @metta-ts/hyperon@experimental

npm install @metta-ts/core continues to resolve the stable release on the
latest tag.

Grounded operation V2 and streaming operations

Grounded operations can now return an owned, pull-based answer stream with
per-answer binding deltas and effects, rather than an eagerly collected array.
The runner exposes MeTTa.registerStreamingOperation and
registerAsyncStreamingOperation: the function returns an iterable (or async
iterable) of answers and the evaluator pulls one at a time, so a consumer such as
once stops the producer instead of draining it. Each answer may bind the call's
argument variables and attach effects applied only on that answer's branch. The
low-level registerGroundedOperationV2 protocol underneath owns and closes every
cursor exactly once, bounds each pull with a finite allowance, and rejects foreign
variables in answers. This is MeTTa TS's implementation of the Minimal MeTTa
document's unimplemented "grounded operations returning bindings" future work; it
has no upstream parity oracle, so its ownership, retention, and binding-delta
behavior are pinned by an executable law suite and a protocol fuzzer.

Cursor streaming through interpreted evaluation

A metta/metta-thread call under a cooperative cursor and interpreted rule
alternatives now stream one answer per pull, so once over a large interpreted
producer performs one step and closes the tail. A single-answer chain keeps its
direct tail-call path; a producer fault after one delivered answer still delivers
that answer first.

Minimal MeTTa runtime foundations

Typed runtime foundations for the minimal machine: scoped variables and canonical
persistent binding frames, binding capture and replay, coherent evaluation
contexts with logical snapshots, explicit Minimal MeTTa control semantics,
resumable search cursors with owned concurrency, and persistent branch worlds with
bounded effect replay.

eval.ts modularization

The core evaluator file was split from about 15,000 lines into a 4,400-line core
plus sixteen leaf-ward modules (machine types, terms, environment, world,
specializer, type views, generator driver, query, concurrent merge, tabling,
cursor constructors, world mutation, matching, fast paths, scheduler glue, and the
eval operation), with the public API preserved through the facade. The change is
mechanical and behavior-preserving.

Correctness: substitution resolves to a fixpoint

A specialized forward chainer emulating backward chaining on propositional
calculus returned an extra spurious proof (GitHub issue #2). Applying a binding
set as a substitution was single-pass, so a variable whose value mentioned another
still-bound variable left that inner variable unresolved; a later scope
restriction then dropped its binding and lost the derived constraint. instantiate
now resolves to a fixpoint. The query returns the single proof that Hyperon 0.2.10
and PeTTa (SWI-Prolog) both produce. The resolution stays bounded and scalable: a
variable chain is followed iteratively (a four-million-link chain resolves rather
than overflowing the stack), a name-to-value index makes a long chain linear rather
than quadratic, binding cycles truncate deterministically, and a shared value DAG
is resolved once by object identity, so a term with an exponential number of paths
resolves in constant time with bounded memory.

Verification

• Full test suite: 147 files pass, 1,849 tests pass, 38 optional live integration
  tests skipped. The Grounded V2 protocol adds an executable law suite, a scripted
  cursor fuzzer, and a disjoint-set binding oracle. The substitution fix adds
  fixpoint and backward-chaining regression tests, each shown to fail on the
  single-pass version.
• Core/ST conformance is unchanged from 1.1.6: 431 passed, 77 established
  failures, 60 manifest expected failures, byte-identical failure set.
• Performance is neutral against 1.1.6 on the nondeterminism benchmark suite;
  the substitution change adds asymptotic headroom (linear chains, DAG sharing)
  with no measured regression.