roundtrip-hebrew-row-codec

The groupant beside tools/roundtrip/src/blocks/site_heb.rs in the Mova repository: what the code computes, what it never does, how to go back from its outputs, its invariants, cost and checks.

Facts

Implementation material. A Mova code model describes this code at one source revision: partial models, finite checks and their stated scope. It is not a reviewed general exposition, and its labels are shown as written, never raised.

Components

component	label	where
approximant	partial	approximant
ast	partial	`syntax.json`
composition	partial	composition
cost	symbolic source derived	invariant-and-cost
descriptant	present	descriptant
invariant	partial	invariant-and-cost
tests	executed bounded checks	tests

Obernants

The typed relations around the forward model: the behaviour complement (what the code never does), the reverse reconstruction (every input compatible with an output) and specialised views.

id	type	label	where
`behavior_complement`	behavior_complement	partial source derived model	behavior_complement
`reverse_reconstruction`	reverse_reconstruction	partial source derived model	reverse_reconstruction
`analysis_precedence`	analysis_precedence	partial source derived model	analysis_precedence
`head_transport`	head_transport	partial source derived model	head_transport
`pair_canonicality`	pair_canonicality	partial source derived model	pair_canonicality
`partial_learning`	partial_learning	partial source derived model	partial_learning

Parts

Model text

descriptant

Description and epoch memory

predict chooses given/exact/bare/compose/segment/unknown analyses, emits ranges and numbered words, derives Ref/MId, uses inner head or next-UPOS then generic offset. diff stores fields or canonical keyed patches; apply sorts keys. Learn tallies analyses/offsets before chunk-cut validation, chooses BTree lexical/offset tie winner, derives affix summaries. write/read serializes core lexicon and recomputes affixes.

Epoch 2026-10-31

Prospective model retained externally before edits. Existing production bytes preserved; only bounded inline tests appended to six owned files. Models distinguish byte reconstruction with residual from semantic understanding and persistent learning. Finite tests and compiled mutants support selected boundaries only. Source inspection identified noninjective projections and explicit admissibility restrictions; no bug repair or performance claim.

Combiner input is this source snapshot, syntax projection, models, costs and retained receipts. Proposed next epoch: target the stated unexecuted branches with bounded witnesses before expanding any claim. AST remains partial, macros opaque; no name resolution or proof.

approximant

Forward relation

Domain: Ten fields per row; unique canonical FEATS/MISC keys for pair identity, '_' empty convention, delimiter-free nonempty analyses for serialization. Finite counts/products/indices, signed offsets avoiding overflow; supplied lexicon only, no global load.

No executable percentage surrogate is supplied: these categorical codec/effect observations have no justified lossy metric. Future caps are 1/2/5/10% only after defining metric, zero behavior and differential reference.

composition

Local composition

Children scope selected internal mechanisms by byte ranges; remaining branches are covered only by the parent partial model. Their costs are included in parent costs, not added again. Shared values and caller order are those named above; sibling fragments are tagged independent where no direct call is named.

catalog

Source-bound typed models

Each section is a partial source-derived relation on the bound snapshot and declared domain; finite checks are falsification evidence, not a universal proof. Observation horizon is one call unless history is explicitly named.

behavior_complement

Given analysis outranks lexicon; unknown token does not acquire a known lemma. External heads clamp but inner heads need not. Noncanonical MISC uses whole field. Lexicon text codec does not preserve arbitrary derived caches or counts.

Domain and held state: Ten fields per row; unique canonical FEATS/MISC keys for pair identity, '_' empty convention, delimiter-free nonempty analyses for serialization. Finite counts/products/indices, signed offsets avoiding overflow; supplied lexicon only, no global load.