Summary

Structural extraction now preserves matching semantics around soft and exact anchors. Inline patterns, grouped patterns, and equivalent definition references have the same match set across nullable paths, alternations, quantifiers, recursion, captures, and grammar fields.

Previously, lowering reduced nullable control flow to consuming and empty outcomes. That lost distinctions such as a pending anchor or the namedness of the previous consumer before control crossed a reference boundary.

Reference output remains opaque. This change preserves structural behavior without exposing captures from an uncollected reference.

Boundary semantics

The compiler now computes a path-sensitive BoundaryRelation for each pattern. Recursive definitions reach a least fixed point per dependency component.

The relation tracks:

• whether the pattern consumed a node
• the first and most recent consumer class
• pending soft or exact anchors
• distinct empty and consuming outcomes

Consecutive anchors tighten their constraint, so an exact anchor can never be loosened by a later soft anchor. Quantifier loops admit only consuming iterations, while ? and * retain their zero-iteration paths.

Lowering maps descriptive outcomes onto an eight-member operational exit universe. Each definition specialization exposes only its reachable ports through a dense local PortId signature. Missing ports mean failure and never fall back to another continuation.

Navigation and grammar fields are carried through a compositional EntryObligation. Empty wrappers and nullable prefixes forward the obligation until the first actual consumer discharges it exactly once. Empty paths and later followers do not inherit the field.

Definitions may now export a boundary-anchor requirement for their caller to satisfy. They remain reusable fragments but are not treated as contextless entry points. Validation reports only anchors that remain without context after definition composition.

Calls and runtime

Ordinary, routed, nullable, and boundary-aware references now use one semantic call operation with:

• caller-owned or callee-owned entry work
• an exact callee port signature
• one continuation per reachable port

The bytecode format is version 12 and uses:

• 0x6 Call1 for one-port calls
• 0x7 CallN for calls with two to eight ports
• 0x8 Return with a callee-local port

The loader verifies call arity, dense return ports, target instruction boundaries, entry ownership, navigation and field contracts, cursor behavior per port, and reserved encoding space.

Runtime frames store only the immutable call-site address. Return(port) pops that frame and resolves (call_site, port) through the call’s continuation map. This preserves cactus-stack restoration and retry behavior without copying a maximum-sized continuation array into ordinary frames.

Generated Rust matchers use the same call-site return model as the bytecode VM.

Validation

Regression coverage now includes:

• inline, grouped, and referenced extraction equivalence
• anonymous trivia that soft anchors may skip and named gaps they must reject
• consecutive exact and soft anchors
• nullable prefixes, suffixes, alternations, and quantifiers
• greedy and lazy retry ordering
• recursive boundary-aware references
• optional, list, text, boolean, record, and variant output
• grammar fields on consuming and empty paths
• VM and generated Rust call behavior
• rejection of malformed multi-port bytecode

A proof-oriented unit test also checks that the operational exit quotient preserves every future anchor observation represented by the descriptive boundary relation.