📐mlab-fileio

The groupant beside mlab/src/fileio.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

Components

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.

Parts

• mlab-fileio-csv-parser
• mlab-fileio-json-table

Model text

descriptant

mlab-fileio

CSV and JSON parsers, inferred table columns and matrix/value serialization without filesystem effects.

Modeling epoch — 2026-10-31

Prospective source read preceded edits. Bounded synthetic checks and selected existing safe tests exercise actual code; paired debug/release results and compiled negative-control assertions are retained externally. Frozen scenario, corpus, external measurement, full-school and training effects are unexecuted. Partial lexical syntax supplies no macro-expanded or semantic proof.

Cross-review refinement — 2026-10-31

Table CSV encodes numeric/date NaN as empty fields, whereas matrix CSV emits NaN text. Logical NaN becomes JSON true by nonzero conversion; numeric nonfinite values become null. Recursive JSON rendering repeatedly copies child text, so cumulative subtree sizes, rather than final output alone, determine work. CRLF/LF ambiguity is limited to record separators; quoted payload retains CR.

approximant

Forward relation

CSV scans Unicode chars after leading BOMs, tracks line/column, decodes quoted fields and validates table width before inference. Numeric inference precedes boolean/date/text and quoted-present forces text. JSON parser preserves object member sequence; table conversion unions first-seen keys, picks first duplicate and fills missing null. Writers emit canonical bytes.

Domain

Valid Rust value/storage invariants and available bounded resources; exact branch and effect observations only.

Executable approximant gap

No useful percentage-error surrogate is established for these categorical, textual or effectful observations. Supported future caps are 1%, 2%, 5%, 10%; no executable approximation or production substitution is claimed.

invariant

Invariants and costs

CSV coordinates are one-based Unicode-scalar columns and decoded line counts. Matrix output uses row order over column-major storage. Parsing allocates values but performs no host/VFS effects.

Failure ordering

CSV syntax errors precede table width errors; quoted unterminated error reports field origin, unexpected postquote reports current location. JSON parser is permissive f64 grammar rather than strict RFC number validator; recursion/allocation unbounded.

Cost

CSV O(B+cells) time/storage. JSON character parsing O(B), error position rescans prefix; key union and per-column row lookups can require O(rows*keys^2). json_text copies child strings at each nesting level: cumulative work is the sum of serialized subtree sizes, bounded by O(depth * output bytes), plus table/cell work. A depth-D singleton-array chain can require O(D^2) copying for O(D) output; recursion depth is unchecked.

composition

Composition

CSV scans Unicode chars after leading BOMs, tracks line/column, decodes quoted fields and validates table width before inference. Numeric inference precedes boolean/date/text and quoted-present forces text. JSON parser preserves object member sequence; table conversion unions first-seen keys, picks first duplicate and fills missing null. Writers emit canonical bytes.

Numeric/date NaN table cells serialize empty CSV; write_matrix emits literal NaN through num_text. Numeric-class nonfinite JSON values become null, while logical-class values use nonzero truth, so logical NaN becomes true. Booleans with missing values become numeric; dates encode day values. Empty table CSV is newline and JSON is [] newline; read_matrix skips at most one nonnumeric header.

csv-parser

Strip leading BOM chars, skip blank LF/CRLF lines; each field records starting line/col, quoted strings decode doubled quotes and embedded LF before delimiter/end handling.

Boundary: Field coordinates are Unicode scalar indices, one-based; each record retains start line and end column.

json-table

Require array of objects, collect keys in first-encounter order, sanitize/dedupe names; each column uses first matching member per row or Null and ordered type tests.

Boundary: Output column order follows first key encounter; all output columns have one element per input object.

Interaction limit

Local fragment relations compose only with valid Value storage, caller error propagation and delegated providers. They do not establish whole-interpreter or host-filesystem guarantees.

models

Typed models — mlab-fileio

Source-derived partial relations; finite checks do not prove universality.

behavior_complement

Quoted numeric CSV fields exclude numeric inference for their column; entirely missing column excludes text inference. J::Num and numeric-class nonfinite JSON serialization excludes numeric literals and becomes null; logical-class values instead use x!=0, so logical NaN serializes true. Successful parse_json excludes trailing nonwhitespace.

reverse_reconstruction

Canonical tables admit many CSV antecedents differing quoting, headers sanitized/deduped and numeric spelling. JSON table preimages include absent versus null and later duplicate keys. Vector JSON cannot reconstruct row versus column orientation or complex imaginary data.

state_invariant

CSV coordinates are one-based Unicode-scalar columns and decoded line counts. Matrix output uses row order over column-major storage. Parsing allocates values but performs no host/VFS effects.

failure_precedence

CSV syntax errors precede table width errors; quoted unterminated error reports field origin, unexpected postquote reports current location. JSON parser is permissive f64 grammar rather than strict RFC number validator; recursion/allocation unbounded.

representation_transport

Numeric/date NaN table cells serialize empty CSV; write_matrix emits literal NaN through num_text. Numeric-class nonfinite JSON values become null, while logical-class values use nonzero truth, so logical NaN becomes true. Booleans with missing values become numeric; dates encode day values. Empty table CSV is newline and JSON is [] newline; read_matrix skips at most one nonnumeric header.

resource_cost

CSV O(B+cells) time/storage. JSON character parsing O(B), error position rescans prefix; key union and per-column row lookups can require O(rows*keys^2). json_text copies child strings at each nesting level: cumulative work is the sum of serialized subtree sizes, bounded by O(depth * output bytes), plus table/cell work. A depth-D singleton-array chain can require O(D^2) copying for O(D) output; recursion depth is unchecked.

tests

Evidence and limitations

Selected named outcomes: fileio::wave30_tests::csv_inference_and_json_loss. Each ran twice debug and twice release with sorted identical nonempty outcomes and numeric exit 0. Source-specific compiled control exercised "fileio::wave30_tests::csv_inference_and_json_loss": compile exit 0, assertion-red exit 101, exact-restored rerun exit 0. External receipts: octopus-wave30-io-2026-10-31 (native-commands.json, positive-receipts.json, controls.json).

Scope: CSV coordinates are one-based Unicode-scalar columns and decoded line counts. Matrix output uses row order over column-major storage. Parsing allocates values but performs no host/VFS effects.

AST projection generated twice to fresh files and compared byte-for-byte. Runtime evidence is finite and does not establish every branch, allocation failure, backend implementation or effect path. No optimization or speed claim.

This page in Markdown → https://md.groupants.elamur.ai/free/en/mova/mlab/src/fileio.rs.groupant/