solverpilot 0.3__tar.gz → 0.4__tar.gz

solverpilot-0.4/BACKEND-CONTRACT-V0_4_0.json

@@ -0,0 +1,51 @@
+{
+  "milestone": "RELEASE_0_4",
+  "notes": [
+    "Backend IDs remain stable across the SolverPilot pre-GitHub cleanup.",
+    "Verification-only adapters are never admitted to the default solving registry.",
+    "Availability of optional public backends remains environment-dependent.",
+    "The legacy optimind Python package is not shipped in SolverPilot 0.1.",
+    "Clarabel has a separate conic capability/conformance surface; it is not an LP/MILP registry entry.",
+    "Unified catalog includes specialized conic/NLP/CP adapters; the legacy LP/QP registry remains compatible.",
+    "Global and PDLP adapters are explicitly selected; existing default registry IDs are unchanged.",
+    "SCIP global bounds are solver-reported numerical evidence, not independent generic global certificates."
+  ],
+  "package_version": "0.4",
+  "production_baseline_policy": {
+    "convex_qp": "osqp-native when available; otherwise scipy-slsqp-qp-bridge for non-proof intents",
+    "learned_lp_performance_routing": false,
+    "lp": "scipy-highs-ds",
+    "milp": "scipy-highs-bridge"
+  },
+  "schema": "solverpilot.backend_contract.v1",
+  "stable_backend_ids": {
+    "base_runtime": [
+      "scipy-highs-ds",
+      "scipy-highs-ipm",
+      "scipy-highs-bridge",
+      "scipy-slsqp-qp-bridge"
+    ],
+    "optional_public": [
+      "highspy-native",
+      "osqp-native",
+      "pyscipopt-native",
+      "nlopt-slsqp-native"
+    ],
+    "verification_only": [
+      "casadi-osqp-bridge",
+      "casadi-highs-bridge",
+      "casadi-cbc-bridge",
+      "bundled-osqp-capi",
+      "bundled-highs-capi"
+    ]
+  },
+  "optional_conic_backends": [
+    "clarabel-native"
+  ],
+  "optional_global_backends": [
+    "scip-global"
+  ],
+  "optional_first_order_backends": [
+    "ortools-pdlp"
+  ]
+}

{solverpilot-0.3 → solverpilot-0.4}/CHANGELOG.md

@@ -1,5 +1,19 @@
 # Changelog
 
+## 0.4 — Conservative certificates and explicit global optimization
+
+- Reject the near-indefinite QP counterexample; require exact represented-data PSD evidence before independent QP bounds.
+- Add a separate matrix nonconvex QP/MIQP IR and bounded factorable SCIP global path, including nonlinear equalities, two-sided rows, general integers, explicit nonsmooth atoms and interval-certified indicators.
+- Add generalized power cone modeling, hashing, Clarabel transport and original-space validation.
+- Reconstruct Clarabel duals and check conservative original-domain SOC/RSOC/PSD bounds; leave unsupported dual membership explicitly unverified.
+- Add isolated CPU PDLP for LP and diagonal convex QP, with canonical duals and wall budgets.
+- Preserve ambiguous primal status for PDLP dual infeasibility; require a checked feasible origin before claiming primal unboundedness.
+- Add optional bounded LP witness recovery and independent rechecking; expose `certificate_recovery` on high-level solve calls.
+- Skip OSQP matrix updates for vector-only edits and expose common reuse observations without claiming factorization reuse.
+- Add shadow policy observations and exact/GPU prerequisite probes, with production learned routing and unqualified integrations disabled.
+- Harden global model snapshots, domains, callback cleanup and numerical result storage; add analytic, exhaustive-reference and adversarial regressions.
+- Preserve historical API/release artifacts; current top-level signatures gain only the documented optional recovery keyword.
+
 ## 0.3 — Numerical trust, sparse modeling and scenario analysis
 
 - Correct Farkas, recession and LP/QP optimality checks using domain-aware bounds and exact operations on represented binary64 coefficients. Bound free QP residuals when a conservative strong-convexity bound is available.

{solverpilot-0.3 → solverpilot-0.4}/KNOWN-LIMITATIONS.md

@@ -1,6 +1,6 @@
-# Known limitations — 0.2
+# Known limitations — 0.4
 
-Version 0.3 adds checked convex atoms, sparse affine lowering, bounded streaming, sensitivity and scenarios. See the current [0.3 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.3.md) for guarantees and remaining scope limits.
+Version 0.4 adds explicit global optimization, generalized power cones, CPU PDLP and conservative certificate recovery. See the [0.4 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md) for exact scope and APIs.
 
 ## Numerical evidence
 
@@ -10,9 +10,9 @@ Per-call ValidationTolerances is supported by execute, solve, solve_production a
 
 ## Conic and nonlinear scope
 
-Optional direct Clarabel supports convex quadratic objectives and affine SOC, rotated SOC, PSD, exponential and power cones. Runtime conformance and original-space validation gate its use. CasADi SuperSCS retains its restricted linear-objective SOC/RSOC scope; its generic PSD and quadratic-conic routes remain disabled. Generalized power cones are not implemented. Conic primal feasibility alone does not imply independently verified global optimality.
+Optional direct Clarabel supports convex quadratic objectives and affine SOC, rotated SOC, PSD, exponential and power cones. Runtime conformance and original-space validation gate its use. CasADi SuperSCS retains its restricted linear-objective SOC/RSOC scope; its generic PSD and quadratic-conic routes remain disabled. Generalized power cones are supported by direct Clarabel. Conservative independent SOC/RSOC/PSD bounds require exact dual membership, objective PSD evidence and a finite residual correction; transcendental cone dual certificates remain unsupported. Conic primal feasibility alone does not imply independently verified global optimality.
 
-Smooth NLP remains optional through CasADi/Ipopt and produces local candidates. Generic global NLP and arbitrary nonsmooth atoms are unsupported. MINLP remains the certified convex binary-discrete scope: general integer, nonconvex/global, nonlinear-equality, two-sided nonlinear and indicator+MINLP compositions remain unsupported. Existing proof flags retain their documented solver-certified scope.
+Smooth NLP remains optional through CasADi/Ipopt and produces local candidates. An explicit SCIP path now accepts bounded factorable global NLP/MINLP, general integers, nonlinear equalities/two-sided rows and supported abs/max/min atoms. Indicator bodies require provable finite interval bounds. Arbitrary Python functions and independent generic global proofs remain unsupported. The existing convex binary OA path retains its previous restrictions and proof semantics.
 
 ## Execution and ownership
 
@@ -20,6 +20,8 @@ Batch accepts canonical LP/QP, uses spawn and requires a script main guard. Time
 
 ReoptimizationSession owns a model clone and serializes update+solve. Session/PersistentSession serialize their public methods, but direct external model/backend mutations require caller coordination. Compiler caches are not shared mutable workspaces for uncoordinated modeling. Reuse is conditional on structure/settings and is not guaranteed faster. Session history can intentionally grow.
 
+PDLP supports only continuous LP/nonnegative diagonal QP and is process-isolated. It is not automatically selected. Exact SCIP and GPU readiness probes do not qualify execution; neither corresponding execution adapter is shipped.
+
 CP-SAT remains process-isolated to avoid known HiGHS ABI collisions. Reference CP and exact TSP/VRP are exhaustive and only suitable for small oracle cases.
 
 ## Modeling and records
@@ -32,4 +34,4 @@ Learned routing remains disabled; OOD indicators are not calibrated probabilitie
 
 Cross-platform support requires exact-artifact qualification on Python 3.12–3.14. Upstream package availability alone is not integration evidence. Synthetic latency/RSS measurements do not establish industrial scalability or absence of long-term leaks. Historical reports retain their original versions and cannot qualify new artifacts.
 
-See [the 0.2 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.2.md) for APIs and examples.
+See [the 0.4 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md) for APIs and examples.

{solverpilot-0.3 → solverpilot-0.4}/MANIFEST.in

@@ -43,3 +43,8 @@ include PUBLIC-API-V0_3_0.json
 include BACKEND-CONTRACT-V0_3_0.json
 include docs/release/SOLVERPILOT-PUBLIC-DOCS-0.3.md
 include docs/release/README-CHECKLIST-0.3.md
+
+include PUBLIC-API-V0_4_0.json
+include BACKEND-CONTRACT-V0_4_0.json
+include docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md
+include docs/release/README-CHECKLIST-0.4.md

{solverpilot-0.3 → solverpilot-0.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: solverpilot
-Version: 0.3
+Version: 0.4
 Summary: Trust-aware optimization runtime for mathematical and constraint programming in Python
 Author: Amirhossein Heidari Rashtabad
 Maintainer: Amirhossein Heidari Rashtabad
@@ -74,11 +74,15 @@ Provides-Extra: minlp
 Requires-Dist: casadi==3.7.2; extra == "minlp"
 Provides-Extra: cp
 Requires-Dist: ortools==9.15.6755; extra == "cp"
+Provides-Extra: pdlp
+Requires-Dist: ortools==9.15.6755; extra == "pdlp"
+Provides-Extra: global
+Requires-Dist: pyscipopt<7,>=6.2.1; extra == "global"
 Dynamic: license-file
 
 # SolverPilot
 
-> **Current version:** `0.3`. Repairs scale-sensitive numerical certificates, preserves sparse model data and solver workspaces, and adds convex atoms, streaming batches, sensitivity, scenarios, robust counterparts and application templates. Historical `0.1.0rc2` correctness/trust hardening remains in place. See the [0.3 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.3.md) for examples and precise guarantees.
+> **Current version:** `0.4`. Repairs near-indefinite QP certificates and adds explicit global QP/MIQP/MINLP, generalized power cones, CPU PDLP, bounded LP witness recovery and reuse observations. Historical `0.1.0rc2` correctness/trust hardening remains in place. See the [0.4 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md) for examples and precise guarantees.
 
 SolverPilot is a **trust-aware optimization runtime and modeling layer for Python**. It combines a small matrix-first solve API for LP/MILP/convex QP with a higher-level semantic modeling system that can compile into conic, smooth nonlinear, certified convex binary MINLP, and constraint-programming execution paths.
 
@@ -111,11 +115,11 @@ SolverPilot has two complementary layers.
 - proof-aware **certified convex binary MINLP** orchestration;
 - integer/Boolean/interval constraint programming with an exhaustive reference backend and optional OR-Tools CP-SAT.
 
-The extended APIs are canonically imported from submodules such as `solverpilot.model`, `solverpilot.conic`, `solverpilot.nlp`, `solverpilot.minlp`, and `solverpilot.cp`. The M30-frozen top-level `solverpilot.__all__` surface remains unchanged in this merge candidate.
+The extended APIs are canonically imported from submodules such as `solverpilot.model`, `solverpilot.conic`, `solverpilot.nlp`, `solverpilot.minlp`, `solverpilot.globalopt`, and `solverpilot.cp`. The M30-frozen top-level `solverpilot.__all__` surface remains unchanged in this release.
 
 ### Verified trust, evaluation, extension, and application layers
 
-The `0.2` release keeps the S2–S10 additive layers and adds correctness/trust hardening discovered during adversarial review of `0.1.0rc1`. The historical `0.0.40rc2` artifacts remain separate provenance and are not overwritten.
+The current release keeps the S2–S10 additive layers and adds correctness/trust hardening discovered during adversarial review of `0.1.0rc1`. The historical `0.0.40rc2` artifacts remain separate provenance and are not overwritten.
 
 - `solverpilot.io`: strict local JSON/CSV ingestion, bounded reads, explicit mapping, content hashes, and source provenance;
 - `solverpilot.evaluation`: direct-run oracle construction and decomposed quality/runtime/failure regret;
@@ -137,12 +141,15 @@ These APIs remain submodule-scoped and do not change the frozen 78-symbol top-le
 | Continuous convex QP | ✅ solve | validated candidate; bridge-specific optimality claims stay conservative |
 | Semantic LP/MILP/QP modeling | ✅ | compiles to canonical core IR |
 | Indicator constraints | ✅ restricted exact bridges | fixed-state simplification or finite-bound-certified Big-M; otherwise fail-closed |
-| SOC | ✅ representation/validation; verified restricted solve path | linear-objective supported verification path only |
-| RSOC | ✅ representation/validation; verified restricted solve path | exact map to SOC in verified path |
-| PSD | ✅ representation/validation | generic PSD solving is not authorized by current conformance evidence |
+| SOC | ✅ direct Clarabel + original validation | conservative independent original-domain bounds when dual evidence is available |
+| RSOC | ✅ direct Clarabel + original validation | reconstructed duals and exact cone membership checks |
+| PSD | ✅ direct Clarabel + original validation | independent bounds require conservative PSD evidence; generic matrices can remain unverified |
+| Exponential / power / generalized power | ✅ direct Clarabel | numerical original-space validation; independent dual certification remains unsupported |
 | Smooth continuous NLP | ✅ optional | local-optimal candidate only after validation/KKT checks; **no global NLP proof** |
 | Convex binary MINLP | ✅ restricted | `globally_proven` is solver-certified inside the certified P8 binary/convex scope; `independently_verified_global` remains false until the full proof chain is independently reconstructed |
-| General integer/nonconvex MINLP | ❌ | fail-closed |
+| Nonconvex QP / MIQP | ✅ explicit `GlobalQuadraticProblem` / SCIP | solver-reported numerical global bound; original candidate validation |
+| General integer/nonconvex MINLP | ✅ explicit bounded factorable SCIP path | nonlinear equality/two-sided rows, abs/max/min and interval-certified indicators; no independent generic global proof |
+| CPU PDLP | ✅ optional LP / diagonal convex QP | isolated worker; explicit selection |
 | Constraint programming | ✅ reference solver | exhaustive proof only within reference state budget |
 | OR-Tools CP-SAT | ✅ optional | exact verified integration target: OR-Tools `9.15.6755` |
 | TSP application | ✅ reference / MILP / restricted CP / heuristics | independent route validation; only exhaustive reference paths issue an independent optimality proof |
@@ -157,7 +164,7 @@ These APIs remain submodule-scoped and do not change the frozen 78-symbol top-le
 - Local-file symlink/path checks are fail-closed for static paths, but they are not a sandbox against a hostile process racing filesystem entries between validation and open.
 - `HistoryStore` is a local plaintext SQLite store. It is opt-in and minimizes persisted solver/problem payloads, but it is not encrypted and caller-supplied metadata can still contain sensitive information.
 - TSP/VRP exact reference algorithms are exponential and intentionally state-budgeted; the routing layer does not yet claim pickup-and-delivery, split delivery, stochastic travel time, live traffic, or real-time redispatch support.
-- The local salvage qualification environment did not contain the exact release build pins or OR-Tools runtime; those remain external release gates rather than locally claimed passes.
+- Exact SCIP and cuOpt GPU execution adapters are not shipped; readiness probes cannot qualify execution. See the [0.4 guide](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md).
 
 ## Requirements
 
@@ -464,6 +471,8 @@ A native patch is selected only when the exact granular capability and persisten
 | HiGHS | `python -m pip install "solverpilot[highs]"` | public `highspy` native adapter |
 | OSQP | `python -m pip install "solverpilot[osqp]"` | public OSQP Python adapter |
 | SCIP | `python -m pip install "solverpilot[scip]"` | PySCIPOpt adapter |
+| Global | `python -m pip install "solverpilot[global]"` | explicit SCIP nonconvex QP/MIQP/bounded MINLP |
+| PDLP | `python -m pip install "solverpilot[pdlp]"` | isolated CPU LP/diagonal convex QP |
 | NLopt | `python -m pip install "solverpilot[nlopt]"` | NLopt SLSQP integration |
 | CasADi | `python -m pip install "solverpilot[casadi]"` | pinned CasADi 3.7.2 verification bridge |
 | Clarabel | `python -m pip install "solverpilot[clarabel]"` | direct convex conic solver, including PSD, exponential and power cones |
@@ -480,18 +489,19 @@ An upstream package being installable is not by itself SolverPilot integration e
 ## What is not supported or not claimed yet
 
 - learned LP performance routing in production;
-- nonconvex continuous QP in the matrix-first QP path;
-- mixed-integer `QuadraticProblem` / general MIQP as a stable matrix-first path;
-- exponential/power/generalized-power cone families;
-- generic verified PSD solving;
-- global optimality for generic NLP;
-- nonsmooth generic NLP atoms such as arbitrary `abs/max/min`;
-- general-integer or nonconvex/global MINLP;
+- nonconvex or integer variables in the existing convex `QuadraticProblem`; use the separate `GlobalQuadraticProblem`;
+- independent verification for every PSD matrix or transcendental cone dual;
+- independent global optimality for arbitrary NLP/MINLP functions;
+- nonsmooth atoms on the smooth Ipopt path; supported abs/max/min are available on the explicit bounded global path;
+- exact SCIP and GPU execution adapters; readiness probes alone do not qualify them;
 - nonlinear equalities/two-sided nonlinear constraints in the current MINLP proof path;
-- indicator+MINLP bridge composition;
+- global indicator bodies without a provable finite interval enclosure;
 - automatic guarantee that repeated solves use native warm starts/factorization reuse;
 - automatic independent certificates for every infeasible/unbounded backend termination;
-- a per-call validation-tolerance argument on high-level `solve()` / `solve_production()`.
+
+Per-call `tolerances=ValidationTolerances(...)` is available on `solve()` and
+`solve_production()`. The [0.4 scope table](docs/release/SOLVERPILOT-PUBLIC-DOCS-0.4.md#scope-against-the-research-checklist)
+maps each research item to its implemented API and remaining boundary.
 
 See [Known Limitations](KNOWN-LIMITATIONS.md).
 
@@ -517,14 +527,17 @@ Extended modeling:
 13. [`13_cp_sat_optional.py`](examples/13_cp_sat_optional.py) — optional OR-Tools CP-SAT
 14. [`14_persistent_session_optional.py`](examples/14_persistent_session_optional.py) — persistent semantic-model session
 
-See [`examples/README.md`](examples/README.md).
+New in 0.4: [global optimization](examples/22_global_optimization_optional.py),
+[generalized power](examples/23_generalized_power_optional.py), and
+[PDLP/certificate recovery](examples/24_pdlp_and_certificates_optional.py).
+See all 24 examples in [`examples/README.md`](examples/README.md).
 
 ## API Reference
 
 - [Stable top-level Public API](docs/api/PUBLIC-API-v1.md)
 - [Extended Modeling API](docs/api/EXTENDED-MODELING-API.md)
-- [Current frozen top-level API snapshot](PUBLIC-API-V0_1_0RC2.json)
-- [Current core backend contract](BACKEND-CONTRACT-V0_1_0RC2.json)
+- [Current frozen top-level API snapshot](PUBLIC-API-V0_4_0.json)
+- [Current core backend contract](BACKEND-CONTRACT-V0_4_0.json)
 - [Track P merge provenance](TRACK-P-MERGE-PROVENANCE.json)
 
 ## Trust and provenance