lp2graph 0.3.0__tar.gz → 0.4.0__tar.gz

lp2graph-0.4.0/CHANGELOG.md

@@ -0,0 +1,380 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/).
+
+## [Unreleased]
+
+## [0.4.0] - 2026-10-04
+
+### Fixed
+
+- **`pulp<4` until the back-end is ported** (issue #65). PuLP 4.0.0 rebuilt its
+  model classes on a Rust core (no Python-constructible `LpAffineExpression` /
+  `LpVariable`, no `PULP_CBC_CMD`, renamed constraints), so a fresh install of
+  the `solver` or `all` extra resolved a PuLP the back-end cannot drive. Both
+  extras now cap it at `<4` (ADR-0008 addendum).
+- **Summed families the referent does not vary with are summed** (issue #60).
+  The grounder looped only over the dummies a term's bindings use and never
+  read `Term.operator_over`, so `\\sum_{i \\in I, r \\in R} dur_r` grounded
+  to `\\sum_r dur_r` and a summed constant `\\sum_{i \\in I} 5` to `5`: the
+  solved value was wrong while text and schema graph were right. Emitter,
+  ground view and grounder now share one reading of an aggregation's loops,
+  `lp2graph.core.scope` (ADR-0016); a binder whose family disagrees with the
+  referent's declared shape (a sum over a subset family, `\\sum_{r \\in
+  \\mathcal{R_k}} x_r` with `x` over `R`) is refused (`UnsupportedModel`)
+  instead of being summed over the whole slot family.
+- **Ground view: summed terms have edges, pair variables are not the
+  diagonal.** The ground view keyed its sum scope by family but resolved
+  bindings by dummy, so every aggregated term contributed no edge (the
+  `assignment` and `mip_2_4_time_indexed` fixtures grounded to graphs with no
+  edges at all), and it resolved bindings into a dict keyed by family, so
+  `y_{i,j}` over `I x I` was wired to `y[j,j]`. Bindings now resolve by slot
+  position, `abs` aggregates over its free indices as in the grounder, and a
+  recurring summand is one edge with a `multiplicity`.
+- **Emitter: binders print what the model says.** `\\sum_{t \\in T}
+  x_{t-1}` was emitted as the unparseable `\\sum_{t-1 \\in T} x_{t-1}`; a
+  summed constant lost its binder (the round trip changed the model);
+  `\\sum_{j \\in I} y_{i,j}` under `\\forall i \\in I` was printed as
+  `\\sum_{i \\in I} y_{i,j}`; a sum over a subset family printed a dummy
+  the summand does not use (`\\sum_{r_k \\in \\mathcal{R_k}} x_r`); and a
+  counting loop re-bound the row's own dummy. The model round trip was intact
+  for the last three, the text was not.
+- **Hybrid view keeps both bindings of a same-family pair.** `offsets` was
+  keyed by family, so `y_{i,j}` kept only `j`; repeated families are now
+  numbered (`I.1`, `I.2`).
+- **A dumped model loads again.** Since the loader began enforcing the JSON
+  Schema (ADR-0013), `loads(f.model_dump_json())` failed for every model with
+  an untagged domain facet, no provenance or no objective: the model dumps
+  those as `null`, which the schema did not admit for `domain_class`,
+  `domain_role`, `provenance` and `objective` (it already did for `lower`,
+  `upper`, `indicator` and `where`). The schema now admits `null` for them
+  (absent or `null` = unset); every previously valid document stays valid.
+
+### Changed
+
+- **A `where`-predicate is a graph edge** (issue #61). The schema and hybrid
+  views add `uses_parameter` (role `where`) from a constraint to the
+  parameter its quantifier predicate reads; the hybrid edge carries the
+  compared value. Formulations that select rows by attribute get different
+  schema graphs, M6 isomorphism results and WL hashes than before (edges are
+  matched on `type|role`); formulations without predicates are unchanged.
+
+- **M3 clustering 50x faster, bit-identical results.** `distance_matrix`
+  reduces every pair to the non-zero coordinates of the row vector (a zero
+  product adds exactly 0.0, so the dense sum is reproduced coordinate for
+  coordinate), and `agglomerative` caches every group pair's average linkage
+  by group id, recomputes only the merged group's distances with the exact
+  member order of the from-scratch loop, and keeps each row's first-minimum
+  column, so a merge costs O(g) instead of a full O(n^2 x members)
+  rescan. Ties still resolve to the first pair in (row, column) order; a
+  property test with forced ties locks the equivalence to the naive
+  definition. One taxonomy induction over the lab corpus went from 84 s to
+  1.5 s.
+
+### Added
+
+- **Binder letters, big-M products, set algebra in binders**
+  (`rewrite-2026.09.4`): `bigop_lone_binder` gives `\\sum_{n}` the family the
+  declared shapes assign to `n` in the row; `distribute_param` distributes a
+  declared parameter over a parenthesised sum exactly (`M \\left(1 -
+  x\\right)` -> `M - M \\cdot x`, numeric pieces become `2 \\cdot M`);
+  `restricted_set_widen` handles set differences/intersections inside binder
+  groups (brace-aware); `declared_product` and `distribute_param` never fire
+  inside a quantifier tail.
+- **Implicit structure the corpus leaves unsaid** (`rewrite-2026.09.3`,
+  measured against the lab corpus after `rewrite-2026.09.2`): a row without
+  any quantifier gets `\\forall` over the families the declared shapes give
+  its free index letters (`implicit_quantifier`, recorded; a letter that would
+  get two families leaves the row untouched); `\\forall f` / `\\forall i, j
+  \\in J` resolve lone letters the same way (`forall_lone_letters`);
+  memberships listed after the algebra without `\\forall` are the
+  quantifier tail (codec); an equality chain `a = b = c` and a same-direction
+  chain with three or more comparators split into consecutive rows, and an
+  interval membership `t \\in [a, b]` into its two bounds (codec, exact);
+  strict `<`/`>` in the algebra relax to `\\le`/`\\ge` as a recorded
+  approximation (restrictions in a tail keep `<`); a parenthesised single
+  symbol `(\\bar{a})_{ik}` is that symbol (`paren_symbol_unwrap`);
+  `\\begin{matrix}`-style wrappers inside a row are dropped
+  (`env_wrapper_strip`); `\\mathbf{\\mathcal{E}}` and `\\in \\mathbf{I}`
+  are the set (`mathbf_set_unwrap`); `\\text{u}` naming a declared or bound
+  symbol is the identifier (`text_ident_symbol`); `restricted_set_widen` also
+  widens function-style sets `H(e)` and set differences/intersections
+  (`S \\setminus T`, `S \\cap T`) to the base family, recorded; unions are
+  refused.
+- **Row conventions the corpus writes** (`rewrite-2026.09.2`, issue #64;
+  lab measurement: 55 of 6,207 accepted rows parsed before this). Codec: the
+  quantifier tail starts at `\qquad` OR the first top-level `\forall`
+  (a dangling comma before it is dropped); trailing `,` `.` `;` `:` on a
+  row, a term or a tail are typography; binder and quantifier clauses accept
+  `i \in \mathcal{I}`, `i \in \mathit{I}`, `i \in I` and `i, j \in I`,
+  and REFUSE by name what used to bind silently or vanish (a subscripted set,
+  a tuple binder, a range, any clause that is neither a set membership, a
+  where-clause nor a restriction). M1b: `text_ident_script` unwraps
+  `\text{l}`/`\mathrm{l}` inside a script; `declared_product` inserts the
+  `\cdot` between a declared parameter and a declared symbol written side by
+  side (coefficient first, `x_{i} c_{i}` -> `c_{i} \cdot x_{i}`), leaving
+  undeclared names for the parser to refuse. `restricted_set_widen` widens a
+  binder or quantifier over a subscripted/superscripted set (`\\mathcal{S}_{i}`)
+  to its base family and records the widening as a rewrite (the same recorded
+  approximation the repository converter makes); Greek index letters
+  (`\\mathit{delta} \\in T`) bind, and a `:` in a quantifier tail separates a
+  restriction clause.
+- **M1b declaration-driven script resolution** (`rewrite-2026.09.1`,
+  issue #63; corpus evidence: 49 + 21 of the 220 papers still failing the
+  Paper-1 promotion after the #52–#57 batch stall on superscripts and
+  label subscripts). The canonical grammar has no superscripts and every
+  subscript position is an index, so the normalizer now decides what a
+  script is from the document itself — the `%@` header (declared names,
+  shapes) and the body's binders/quantifiers — and rewrites bijectively:
+  `bare_sub_brace`/`bare_sup_brace` brace unbraced scripts against the
+  declarations (`B_u \cdot w_u` -> `B_{u} \cdot w_{u}`, a declared plain
+  name such as `Z_1` stays); `superscript_index` moves bound-letter
+  superscripts into the subscript (`x_{i}^{k}` -> `x_{i, k}`,
+  `p_{n}^{t + 1}` -> `p_{n, t + 1}`, MathML-spaced `^{i j}` -> two
+  indices); `superscript_label` folds label superscripts into plain names
+  (`t_{i}^{arr}` -> `t_arr_{i}`, `v_{i}^{c}` -> `v_c_{i}`,
+  `\mathit{tau}_{k}^{de}` -> `tau_de_{k}`, `Y_{i,s}^{1}` -> `Y_1_{i,s}`,
+  `q^{*}` -> `q_star`, `d^{+}`/`d^{-}` -> `d_plus`/`d_minus`); `label_subscript` folds label subscripts
+  (`h_{min}` -> `h_min`, `Z_{1}` -> `Z_1` when `Z` has no declared shape;
+  a numeric subscript on a shaped symbol stays a fixed-element reference).
+  Nested or delimited scripts are left untouched for the parser to refuse
+  by name, glued bound letters split into indices (`x_{ij}` -> `x_{i, j}`),
+  and a document without any `%@` declaration is left untouched by all
+  five rules (nothing to resolve against). `RewriteRule` gained `ctx_replacement` (rules that read a
+  `DocContext`), and a rule that declines records no rewrite. The
+  declaration sidecar must declare the folded spellings.
+
+### Fixed
+
+- **Codec accepted undeclared identifiers silently** (issue #62): an
+  undeclared referent fell through `_SymTab.kind()` as a `literal` term
+  and a bare coefficient name was returned without checking it is a
+  declared parameter, so unbraced subscripts (`B_u \cdot w_u`) ingested
+  `ok=True` with `ref="w_u"`, `coefficient="B_u"` and did not round-trip
+  (surfaced by the lab's re-promoted `trc.2014.06.003`). Both paths now
+  refuse by name, pointing at the declaration and the braced form.
+
+- **Grammar holes #52–#57 closed deterministically** (issues filed from the
+  Paper-1 corpus promotion sprint; every fix is exact-or-refused-by-name,
+  nothing is dropped silently):
+  - *Codec parser* (`codec.latex`): `\leq`/`\geq`/`\leqslant`/`\geqslant`
+    accepted as row comparators with token guards (previously `\leq` was read
+    as `\le` plus a stray `q`, and `\left|` on a constraint LHS was read as
+    `\le` — both silently corrupted the relation split); chained
+    same-direction inequality rows `l \le e \le u` split into two constraints
+    `<name>_lo`/`<name>_up` inheriting the row's quantifiers (#53;
+    mixed-direction, equality, and 3+-comparator chains refused by name);
+    subscripted coefficients `w_{e} \cdot x_{e}` resolve to the bare
+    parameter exactly when the written indices match what grounding uses —
+    the referent's bindings, or the unique in-scope binder/quantifier index
+    per declared shape slot (#52); numeric `\frac{a}{b}` with a terminating
+    decimal folds into the numeric coefficient, every other `\frac` is
+    refused by name (#57); constant subscripts (`t_{0}`) get a named
+    diagnostic pointing at the missing shape declaration — with the family
+    declared they were already legal element references (#54); trailing
+    `^{...}` or juxtaposed factors after a referent's subscript are refused
+    by name instead of being **silently dropped** (previously
+    `p_{e} x_{e}` parsed with `x` discarded and `x_{e}^{k}` lost its
+    superscript index).
+  - *M1b normalizer* (`mining.ingest.latex_normalizer`): `sum_merge`
+    (consecutive `\sum` operators merge into one multi-binder `\sum`,
+    order-preserving, #56); `overset_accent`/`underset_accent` (accent-shaped
+    `\overset`/`\underset` pairs collapse to the standard accent command
+    before `overset_base` can drop the accent); `accent_ident` and
+    `prime_ident` (decorated identifiers rename bijectively to plain `\w+`
+    names — `\hat{tc}` → `tc_hat`, `t'`/`k^{'}` → `tp`/`kp` — valid in every
+    grammar position, #55). Rule-table version bumped to `rewrite-2026.08.1`.
+
+- **`lp2graph.validation`** — end-to-end validation of (LLM-)generated
+  LP/MILP artifacts. `validate_text` / `validate_path` / `validate_formulation`
+  accept raw text, bytes, files, or parsed models in any supported format
+  (canonical JSON, LaTeX, LP, MPS, GAMS, AMPL, JuMP) and return a structured
+  `ValidationReport` instead of raising: faulty-input detectors and repairs
+  (markdown fences, unicode look-alikes, truncation, NUL/encoding damage),
+  format sniffing with parse fallbacks, the semantic invariants, structural
+  detectors (completeness, coherence, duplicate/unused symbols, constant
+  constraints, bound conflicts), and an optional grounding smoke check on a
+  synthesized all-ones instance (CBC/HiGHS/Gurobi; skipped gracefully without
+  pulp). Reports are deterministic and stamped with a `pipeline_version`.
+- **`lp2graph validate` CLI** upgraded to run that pipeline on any supported
+  model file (previously canonical JSON only): `--fmt`, `--json`, `--no-solve`,
+  `--instance`, `--solver`, `--time-limit`; exit code 0 unless the verdict is
+  `invalid`. New docs page `docs/validation.md`.
+
+- **M1b big-operator wrapper rules** (`mining.ingest.latex_normalizer`) — four
+  new rewrite rules driven by Paper-1 corpus evidence (3,668 of 8,957 Tier-2
+  MathML-derived formulas): `underset_bigop` (`\underset{X}{\sum}` →
+  `\sum_{X}`, likewise `\prod`/`\min`/`\max`/`\int`/`\bigcup`/`\bigcap`),
+  `mathop_unwrap`, `underbrace_unwrap`, and `overset_base`. Rule-table version
+  bumped to `rewrite-2026.07.0`.
+- **`lp2graph.interop`** — functional code ⇄ graph interfaces for the common
+  modeling languages, replacing the historic stubs. Importers build
+  coefficient-faithful flat `Formulation`s from **gurobipy** models
+  (`from_gurobipy`), **PuLP** problems (`from_pulp`), **Pyomo** concrete
+  models (`from_pyomo`, full linear term recovery via `standard_repn`;
+  ranged constraints split), and **LP / MPS / GAMS / AMPL / JuMP** text
+  (`from_lp_string` / `from_mps_string` / `from_gams` / `from_ampl` /
+  `from_jump`). Exporters emit every one of those targets from any
+  formulation (`to_gurobipy[_code]`, `to_pulp[_code]`, `to_pyomo[_code]`,
+  `to_lp_string`, `to_mps_string`, `to_gams`, `to_ampl`, `to_jump`) — flat
+  models directly, template-level models through the PuLP grounder with an
+  `Instance`. All emitters are deterministic fixpoints; unsupported
+  constructs raise `InteropError` instead of being dropped. Verified by a
+  `code → graph → code` round-trip matrix (`tests/interop/`, 100 tests)
+  that solves every path against hand-verified optima with CBC, HiGHS, and
+  Gurobi, including cross-reads of Gurobi-written `.lp`/`.mps` files.
+- **`lp2graph convert IN OUT`** — CLI conversion between modeling languages
+  through the canonical graph, routed by file extension
+  (`.json/.tex/.lp/.mps/.gms/.mod/.jl`, plus `.py` solver scripts via
+  `--python-api {gurobipy,pulp,pyomo}`).
+- **`mining.ingest`** now routes `.gms/.mod/.jl` to the real interop parsers
+  (previously honest stubs) and gained `.lp`/`.mps` support (`import_lp`,
+  `import_mps`); parse problems surface as structured `stage="parse"`
+  failures.
+
+- **`metrics.model_completeness`** — the second model-level well-formedness
+  indicator described in *LP Mining with LP2Graph* (objective declared together
+  with ≥1 variable and ≥1 constraint), companion to `model_coherence`. Now
+  computed and wired into the Level-M structural feature document
+  (`mining.cluster.taxonomy.model_feature_document`) as a `complete:{0,1}`
+  feature, closing a paper↔code gap where the paper named both indicators but
+  only coherence was implemented.
+- **LP mining extensions** (`lp2graph.mining`, issues #38–#43) — six
+  deterministic modules implementing the *LP Mining with LP2Graph* method on
+  top of the core library. Every frozen resource is versioned in
+  `lp2graph.mining.versions` and stamped into emitted records.
+  - **M1 `mining.ingest`** — heterogeneous ingestion front-end: a Pyomo
+    importer (`from_pyomo`), a versioned non-canonical LaTeX normalizer with
+    source-span provenance (`normalize_latex` / `ingest_latex`), and an
+    extension dispatcher (`ingest`) that reports failures as structured
+    `IngestionResult`s rather than dropping them.
+  - **M2 `mining.homologize`** — lexical homologizer (tokenize / lemmatize /
+    versioned stop-list / frozen domain thesaurus + optional WordNet) and a
+    TF-IDF `ConceptVectorizer` over a sorted, diffable `Vocabulary`, plus the
+    type signature `τ(s)` and the `Entity` model for levels V/C/M.
+  - **M3 `mining.cluster`** — the cluster-and-name operator `CN` (deterministic
+    average-linkage default, `fixed_k`+silhouette, optional HDBSCAN), the
+    bottom-up Level V→C→M taxonomy `induce`, and `stability_report`
+    (silhouette + bootstrap ARI + sensitivity).
+  - **M4 `mining.label`** — two-stage labeling service: rule layer + calibrated
+    one-vs-rest `LinearSVM`, a confidence-gated closed loop with a versioned,
+    replayable `LabelStore`, rule promotion, retraining, and gold-set
+    guardrails (drift, per-class P/R, Cohen's κ, rollback flag).
+  - **M5 `mining.corpusmgr`** — provenance records, a regeneration `CorpusManifest`
+    (frozen search date + queries), deterministic dedup (schema-graph hash or
+    bibliographic key), and reproducible representative selection.
+  - **M6 `mining.isomorphism`** — per-cluster schema-graph isomorphism rate via
+    the NetworkX export (`isomorphism_report`).
+
+### Fixed
+
+- **`interop.from_pulp` crashed on unnamed constraints** (`prob += expr <= rhs`
+  without a name label): the constraint's ``name`` is ``None`` and the
+  name-sanitizer raised ``TypeError``. Anonymous constraints now take the
+  deterministic ``c1``/``c2``/... fallback names.
+
+### Changed
+
+- **PuLP 4.0 forward-compatibility** (`lp2graph.solve` + `lp2graph.interop`):
+  migrated off the APIs PuLP 4.0 removes — variables are built with
+  `prob.add_variable(...)`, constraints counted with `prob.numConstraints()`,
+  and the default CBC solver is created by the new `solve.default_solver()`
+  (`COIN_CMD` with a bundled-CBC fallback) instead of the deprecated
+  `PULP_CBC_CMD`. `interop.grounded_from_pulp` lists constraints
+  version-agnostically, `to_pulp_code` emits scripts using the sanctioned
+  APIs, and the M1 ingest reader reports non-UTF-8 source files as
+  structured read-stage failures (ADR-0009) instead of raising.
+
+- **CI now gates formatting:** added a `ruff format --check src tests` step to
+  `ci.yml` (previously only `ruff check` ran, so format drift could land
+  undetected). Bumped the `ruff-pre-commit` pin v0.4.10 → v0.15.12 and the
+  `dev` extra to `ruff>=0.15.12,<0.16` so local, pre-commit, and CI agree;
+  reformatted 21 pre-existing files to the current ruff style.
+
+### Docs
+
+- Added `docs/STACK.md` (software-stack reference) and ADR-0006 (determinism as
+  a hard requirement) and ADR-0007 (optional dependencies are lazily imported),
+  wired into the MkDocs nav.
+
+## [0.3.0] - 2026-06-01
+
+### Added
+
+- **Deterministic LaTeX ⇄ graph codec** (`lp2graph.codec`): paper-style
+  LaTeX (`\mathcal` index sets, `\sum`, `\forall`, big-M) round-trips with
+  the canonical model with no LLM in the loop.
+  `to_canonical_latex` / `from_canonical_latex`, plus
+  `canonical_normal_form` for round-trip comparison. The LaTeX
+  serialization is a tested fixed point (`to(from(to(f))) == to(f)`); the
+  solvable content round-trips exactly.
+- **Real grounding solver back-end** (`lp2graph.solve`): replaces the v0.1
+  Pyomo stub. Grounds a formulation with an `Instance` (cardinalities +
+  parameter values) into a concrete `pulp.LpProblem` and solves it
+  (CBC / HiGHS / Gurobi). Covers the linear core including big-M and PESP
+  modulo; boundary-degenerate constraint instances are correctly omitted.
+- **Deterministic graph → natural-language describer** (`lp2graph.nl`):
+  generates a Markdown problem description (sets, data, decisions,
+  per-constraint sentences, objective) with parameter **data tables** when
+  an instance is supplied.
+- **End-to-end validation suite** (`corpus/validation/codec_pipeline/`):
+  runs the full JSON→LaTeX→parse→ground→solve loop and checks codec
+  round-trip, pipeline-vs-direct equality, cross-solver agreement, and
+  match to independently established optima (assignment 13, big-M ordering
+  30, fixed-sequence 18, PESP 1, time-indexed 4). Records paper anchors
+  (timtab1 = 764772, marcotallone = 3913.47).
+- New formulations: `assignment.json`, `pesp_solvable.json`.
+- CLI: `lp2graph latex | parse | describe | solve`.
+- New tests: `test_codec.py`, `test_solve.py`, `test_describe.py`.
+- Optional `solver` extra (`pulp`, `highspy`).
+
+## [0.2.0] - 2026-05-20
+
+### Changed
+
+- **BREAKING:** the import package is renamed `optgraph` → `lp2graph`, so it
+  now matches the distribution name. Update imports
+  (`from optgraph import …` → `from lp2graph import …`) and the console
+  script (`optgraph …` → `lp2graph …`). No behaviour changed; this is a
+  pure rename. The canonical `schema_version` is unaffected (still `0.1.0`).
+
+## [0.1.0] - 2026-05-03
+
+### Added
+
+- Canonical JSON schema (`schema/canonical.schema.json`) with index
+  families, parameters, variable templates, constraint templates,
+  first-class objectives, and term-level refs/bindings/role/sign
+  semantics.
+- Pydantic v2 model (`lp2graph.core.model`) mirroring the schema.
+- Two-phase validator (JSON Schema + semantic invariants).
+- Three view derivations: `lp2graph.views.schema`, `.hybrid`, `.ground`.
+- Internal typed graph (`lp2graph.core.graph`) — library-agnostic.
+- Structural metrics: `node_counts_by_class`, `edge_density`,
+  `constraint_variable_ratio`, `minimal_size`, `model_coherence`,
+  `graph_diameter`.
+- Presence flags: `has_big_m`, `has_integer_vars`,
+  `has_modulo_offset`, `has_soft_slack`,
+  `has_aggregation_operator`.
+- Constraint classification (heuristic, regex-based; ported from the
+  source repo with attribution).
+- SVG renderer with the design-context palette and typography.
+- Static interactive viewer (`viewer/index.html`).
+- Export adapters: NetworkX (full), PyG (HeteroData), DGL, LaTeX,
+  Pyomo (skeleton; bodies are stubs in v0.1).
+- CLI: `lp2graph validate | view | render | metrics | export`.
+- Initial catalog: 5 constraint-focused and 3 objective-focused
+  formulations across LP and MILP families.
+- 36 tests covering schema validation, view derivations, metrics,
+  rendering, and exports.
+- CI for Linux + macOS, Python 3.11/3.12/3.13.
+- Docs: data-model, views, metrics, add-a-formulation, design context,
+  extraction report; ADRs 0001-0005.
+
+### Origin
+
+Bootstrapped from
+[`joernmht/raiLPminerExperimentation`](https://github.com/joernmht/raiLPminerExperimentation)
+(MIT). See `docs/extraction-report.md` for the manifest.

{lp2graph-0.3.0 → lp2graph-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: lp2graph
-Version: 0.3.0
+Version: 0.4.0
 Summary: Typed-graph representation of LP, MIP, and MILP formulations.
 Project-URL: Homepage, https://github.com/joernmht/lp2graph
 Project-URL: Documentation, https://github.com/joernmht/lp2graph#readme
@@ -27,7 +27,7 @@ Requires-Dist: hdbscan>=0.8; extra == 'all'
 Requires-Dist: highspy>=1.7; extra == 'all'
 Requires-Dist: networkx>=3.0; extra == 'all'
 Requires-Dist: nltk>=3.8; extra == 'all'
-Requires-Dist: pulp>=2.8; extra == 'all'
+Requires-Dist: pulp<4,>=2.8; extra == 'all'
 Requires-Dist: pyomo>=6.7; extra == 'all'
 Requires-Dist: torch-geometric>=2.4; extra == 'all'
 Requires-Dist: torch>=2.0; extra == 'all'
@@ -40,11 +40,13 @@ Requires-Dist: pytest-cov>=4.0; extra == 'dev'
 Requires-Dist: pytest>=8.0; extra == 'dev'
 Requires-Dist: ruff<0.16,>=0.15.12; extra == 'dev'
 Provides-Extra: dgl
-Requires-Dist: dgl>=2.0; extra == 'dgl'
+Requires-Dist: dgl>=2.0; (python_version < '3.13') and extra == 'dgl'
 Requires-Dist: torch>=2.0; extra == 'dgl'
 Provides-Extra: docs
 Requires-Dist: mkdocs-material>=9.5; extra == 'docs'
 Requires-Dist: mkdocstrings[python]>=0.25; extra == 'docs'
+Provides-Extra: gurobi
+Requires-Dist: gurobipy>=11.0; extra == 'gurobi'
 Provides-Extra: mining
 Requires-Dist: hdbscan>=0.8; extra == 'mining'
 Requires-Dist: networkx>=3.0; extra == 'mining'
@@ -59,7 +61,7 @@ Provides-Extra: pyomo
 Requires-Dist: pyomo>=6.7; extra == 'pyomo'
 Provides-Extra: solver
 Requires-Dist: highspy>=1.7; extra == 'solver'
-Requires-Dist: pulp>=2.8; extra == 'solver'
+Requires-Dist: pulp<4,>=2.8; extra == 'solver'
 Description-Content-Type: text/markdown
 
 # lp2graph
@@ -169,6 +171,18 @@ A **formulation** is a JSON document validated against
 *constraint templates* with quantifiers and bindings, and an optional
 *objective* with first-class terms.
 
+`load()`/`loads()` enforce that schema on every document before the typed
+model sees it, so a file that violates the published contract is refused with
+spec-grounded messages naming the source. The schema ships inside the package;
+read it with the accessor rather than by path, since its location differs
+between a wheel and a source checkout:
+
+```python
+import lp2graph
+schema = lp2graph.canonical_schema()        # parsed dict, cached
+path = lp2graph.canonical_schema_path()     # on-disk location
+```
+
 The **schema view** exposes templates and indices — the topology of the
 problem. The **hybrid view** adds per-term offset, sign, and modulo
 labels. The **ground view** materializes every instance at given index
@@ -199,6 +213,12 @@ are visible.
 See [`CONTRIBUTING.md`](CONTRIBUTING.md). Bugs, new formulations, design
 decisions, and good-first-issues all have dedicated templates.
 
+## AI disclaimer
+
+This repository was developed with substantial assistance from AI coding tools
+(primarily Anthropic's Claude). Code, documentation and results have been
+reviewed by the author, who takes full responsibility for the content.
+
 ## License
 
 Apache 2.0. Portions adapted from

{lp2graph-0.3.0 → lp2graph-0.4.0}/README.md

@@ -105,6 +105,18 @@ A **formulation** is a JSON document validated against
 *constraint templates* with quantifiers and bindings, and an optional
 *objective* with first-class terms.
 
+`load()`/`loads()` enforce that schema on every document before the typed
+model sees it, so a file that violates the published contract is refused with
+spec-grounded messages naming the source. The schema ships inside the package;
+read it with the accessor rather than by path, since its location differs
+between a wheel and a source checkout:
+
+```python
+import lp2graph
+schema = lp2graph.canonical_schema()        # parsed dict, cached
+path = lp2graph.canonical_schema_path()     # on-disk location
+```
+
 The **schema view** exposes templates and indices — the topology of the
 problem. The **hybrid view** adds per-term offset, sign, and modulo
 labels. The **ground view** materializes every instance at given index
@@ -135,6 +147,12 @@ are visible.
 See [`CONTRIBUTING.md`](CONTRIBUTING.md). Bugs, new formulations, design
 decisions, and good-first-issues all have dedicated templates.
 
+## AI disclaimer
+
+This repository was developed with substantial assistance from AI coding tools
+(primarily Anthropic's Claude). Code, documentation and results have been
+reviewed by the author, who takes full responsibility for the content.
+
 ## License
 
 Apache 2.0. Portions adapted from

lp2graph-0.4.0/docs/DECISIONS.md

@@ -0,0 +1,23 @@
+# Decision & Open-Points Register — lp2graph (library)
+
+Running log of **decisions**, **open questions** and **notes** for this project.
+Append-only; newest entries at the bottom. Managed by the `transcript` skill via
+`register.py`. The central aggregate of all open points across projects lives at
+[`~/DECISIONS.md`](/home/joern/DECISIONS.md).
+
+**Lifecycle:** `open` → `decided` → (if architectural) graduate into a numbered
+**ADR** in `docs/adr/`. ADRs stay immutable; this register is where things churn.
+
+Entry fields: *Type* (decision/open-question/note), *Status*, *Date*, *Source*
+(meeting/transcript), *Owner*, *Summary*, optional *Rationale* and *Next*.
+
+---
+
+### D-2026-0619-01 · Periodic model fails to solve despite coherence handling
+- **Type:** open-question
+- **Status:** open
+- **Date:** 2026-06-19
+- **Source:** meeting 2026-06-18 (~/meetings/2026-06-18-paper1-status-review.md)
+- **Owner:** JM
+- **Summary:** A periodic formulation does not solve although Paper-1's coherence handling should cover it ('I don't know why this isn't right; look into this'). Investigate in the lp2graph grounding/solver path.
+