Elamur Groupantstreestore-frag
The groupant beside tools/treestore/src/frag.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 |
| Descriptant | present | Descriptant |
| Invariant | partial | Invariant |
| 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.
| relation | type | label | where |
|---|---|---|---|
| Behavior complement | behavior complement | partial source derived model | Behavior complement |
| Reverse reconstruction | reverse reconstruction | partial source derived model | Reverse reconstruction |
Model text
Descriptant
The meta-dictionary of typical structures: frequent runs of token shapes (tag tuple, relation, relative head) learned by pair merging (Gage 1994; Re-Pair, Larsson & Moffat 2000). With relative heads, a run whose heads point inside it is a tree fragment. Epoch 2026-11-06: measured gain over single shapes is small (see approximant).
Approximant
learn: repeat n times: take the most frequent adjacent pair (ties: smallest pair) with count ≥ min, give it a new symbol, replace its occurrences left to right; counts are updated incrementally (a lazy max-heap). encode: merge the lowest-rank pair present until none. expand: symbol → shapes. Measured on shard 00050 (structure stream + escapes): shape ids alone 2.04 MB (context coder), 1024 fragments 2.01 MB (−1.5 %), 4096 fragments 2.12 MB, 16384 fragments 2.30 MB; under zstd only: shapes 2.49 MB, 4096 fragments 2.44 MB. The gain of the structure coding comes from the joint shape symbol (split columns 3.23–3.50 MB → shapes 2.04 MB).
Invariant
expand ∘ encode = id on every shape sequence (also unseen ones; symbols below base expand to themselves). learn is
deterministic. Cost: learn O(merges · affected sentences · length); encode O(len · merges applied).
Composition
model::train learns it from the sample; codec writes encode(shapes) to the shape stream and expands while decoding
until the sentence's token count is reached.
Behavior complement
Does not learn non-contiguous subtrees (only runs in word order), does not use the context coder's statistics, and does not prune fragments that do not pay off.
Reverse reconstruction
Exact: a fragment sequence has one expansion; a shape sequence has one encoding for fixed merges.
Tests
frag::tests::learn_encode_expand (first merge = most frequent pair, round trip, unseen sequence, determinism, the
negative control that an unknown symbol does not expand).
Provenance
This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/tools/treestore/src/frag.rs.groupant/