Work is on codex/solver-parity in the independent checkout at implementation/gecode. The original solver checkout, branch, uncommitted source and benchmark binaries remain untouched. The enhanced source/test/input snapshot is e10562fd9; source-snapshot.json records 2,963 input hashes. All 2,963 original files matched when checked after integration.

This ledger covers the implemented slices of the ongoing W0–W14 roadmap. The validation record identifies the exact source used for each completed gate. They provide numerical LP/MILP modeling and workflows, persistent numerical sessions, typed native globals, weighted feasibility repair, C/Python bindings, an explicit bounded integer compiler to native Gecode search, ranked solution pools, checked integer LP deductions, exact integer presolve/reconstruction, interruption-safe native frontier bounds, bounded reliability branching, verified root covers, continuous weighted-square QP and an owning FlatZinc capture/compiler with a separate command-line solver. Commercial-solver parity and the larger native search improvements remain incomplete. The algorithm checkpoint adds bounded native binary capacity optimization after the ninth checkpoint. Further roadmap implementation is on hold. Each report identifies its own tested source, artifacts and prior-version comparison.

Implemented in this batch

Area	Delivered behavior	Where to start
Native binary capacity optimization	Bounded exact dynamic programming supplies a checked witness preference and objective bound for one-row Binary capacity models; other search defaults are preserved	Algorithm checkpoint
Isolation and dependencies	Independent clone/branch, separate agent worktrees/build directories, preserved enhanced baseline, pinned HiGHS source	source-snapshot.json, dependencies.json
Sparse owning model	Five explicit variable domains; ranged linear rows; min/max/offsets; canonical sparse terms; stable model-aware handles, tombstones, revisions and snapshots	gecode/optimize/model.hpp
Bulk construction	Atomic variable/row batches and explicit CSR input, strong allocation-failure guarantee, geometric growth and no dense matrix allocation	Bulk API
Result and budget contract	Status separate from incumbent; backend/version attribution; numerical guarantee; missing-bound semantics; normalized/native gaps; shared time/cancellation and node budget primitives	gecode/optimize/result.hpp
Numerical backend	Optional HiGHS LP/MILP route, explicit unsupported classes/guarantees/scales, deadline-aware candidate capture, one worker per solve	gecode/optimize/solve.hpp
Independent checking	Structural trust boundary; original bounds, rows, integrality, objective and logical indicators; original-slot solution mapping	gecode/optimize/validate.hpp
Starts and edits	Numerical partial/complete starts; native exact complete starts with live indicator gate completion and strict improvement search; model edits and historical result ownership	SolveOptions::primal_start, Model
Persistent sessions	Compatible cost/bound/RHS edits retain LP bases; structural/type/owner changes reload; MIP witnesses are revalidated before reuse; fresh options, budgets and callback lifetime per solve	Sessions
Explicit LP basis starts	Immutable original-model basis records, precise compatibility rejection and accepted/repaired submission reporting, with C/Python ownership	Basis API, Bindings
Serial scenario batches	Sparse coefficient, offset, variable-bound and row-side overrides; independent historical results, shared time/cancellation and optional private HiGHS session reuse, with C/Python ownership	Scenarios, Bindings
Numerical LP evidence	Explicit private auxiliary solves recover checked original feasible bases/improving rays or signed Farkas contradictions; separate availability and raw stage history, with C/Python ownership	LP evidence, Bindings
Selected-basis LP sensitivity	Numerical objective-coefficient and equality-RHS intervals, private checked basis factorization without optimization runs, owning results and C/Python lifetime/cancellation behavior	Sensitivity, Bindings
Local Python wheel	Self-contained macOS arm64 C ABI, automatic package-relative loading, pinned notices and RECORD checks; isolated Python 3.12/3.14 installation tests	Wheels
Continuous LP observations	Owning original activities, two-sided slacks, duals, reduced costs and basis export with independent numerical KKT checking and C/Python access	LP observations
Native integer route	Explicit Backend::Native, actual Space/IntVarArray and branch-and-bound, native reified indicators, exact original-integer witness checking within documented activity limits	Native bridge
Native global registry	Typed all-different, element, table, cumulative, circuit and Regular records; aliases/index bases retained; independent predicates and native compiler; automatic native dispatch for active globals	Globals
Sparse checked integer LP	Binary and bounded-integer CSR storage, exact dual bounds and interval reductions, sibling-bound restoration and explicit native hybrid API; dense compatibility retained	Sparse LP, Native LP
Native search frontier	Explicit best-bound or depth-first coordinator; exact interrupted global bounds, retained incumbents and bounded live regions, optional checked LP and opt-in bounded binary reliability probes	Native search
Native incumbent neighborhood	Explicit C++ BinaryHamming attempt after a checked incumbent, original-model improvement validation and shared bounded frontier accounting	Neighborhoods
Verified root covers	Budgeted binary cover separation, independent exact cut validation and explicit handoff to native LP/frontier search; separate cut and LP accounting	Root covers
Continuous quadratic models	Distinct owning weighted-affine-square model, bounded convex min/concave max through HiGHS QP, original primal/KKT checks and outward tangent bound; C/Python bindings	Quadratic API
FlatZinc pipeline	Owning capture, strict whole-model integer/Boolean compiler, supported implications/reifications, all-different/element, table/holey domains, explicit-offset circuit fixed-data cumulative and literal deterministic Regular; independent raw-source witness validation and explicit fzn-gecode-optimize CLI	Compiler, Driver
Experimental MiniZinc route	Opt-in separate solver identity and dedicated library, explicit millisecond protocol, relocatable build/install registration and real pinned compiler/output checks	MiniZinc
Exact presolve/postsolve	Signed interval implications, fixed substitution, redundant rows and immutable original/reduced mappings; explicit utility with exact historical witness reconstruction	Presolve
Solution pools	Ranked representatives over finite discrete projections, continuous recourse, original witness validation and separate RequestedLimit/Exhausted evidence	Pools
Feasibility repair	Explicit weighted L1 row/bound relaxation on a distinct private model; optional ordered original-objective refinement; original residual reporting and separate proof/completion fields	Repair
C and Python	Versioned C99 owning-handle boundary, immutable results, cancellation, sessions and model I/O; Python ctypes package; installed C consumer and backend-enabled/disabled conformance	Bindings
Conflict diagnostics	Numerical deletion-minimal groups for rows, bounds, integrality and domains; original attribution and optional deletion witnesses; interruption cannot promote an incomplete conflict	Diagnostics
Fast regression gate	Required native CP and numerical cases, plus sessions, diagnostics and exact native bridge; a 28-second whole-command budget and explicit incomplete-case failures	Fast gate
Bounded indicators	Conservative domain-derived M, separate binary inactivity gate, preserved original logic, atomic posting, mutation/domain guards and checked removal	Constraint helpers
Boolean helpers	AND/OR equality, empty/single/multiple inputs, aliases and duplicates; atomic posting	Constraint helpers
Multiple objectives	Ordered linear stages, explicit degradation, one outer deadline, retained original-model feasible result, no false completion after interrupted stages	Workflows
Model exchange	Strict numerical LP/free-MPS readers; 17-digit output; checked round trips; atomic replacement; explicit unsupported dialects	I/O
Distribution and CLI	Standalone or opt-in native CMake component; installed package aliases; umbrella header; JSON file solver; isolated exact-check benchmark harness	Build and API guide
Verification and CI	Independent component tests, full-source sanitizers, cross-feature integration, native regressions, installed consumers, five existing MIPLIB models; added platform CI workflow	Validation

The backend's search, presolve, cuts and heuristics in this new route are HiGHS capabilities. Native Gecode CP and the checked sparse relaxation remain available through their original interfaces. The explicit native hybrid adds certificate-checked LP deductions to native propagation; its results identify both engines. Ordinary numerical solves continue to use HiGHS search and ordinary native routing is unchanged.

Roadmap status and deferred completion gates

These gates describe unfinished scope; they are not authorization to start a new batch after the current integration checkpoint.

Work package	Status after this batch	Next completion gate
W0 reproducible foundation	Source isolation/pin, correctness harness and bounded fast regression gate implemented; comprehensive performance baseline still outstanding	Freeze cold/warm tracks, machine/environment accounting and untouched holdout inputs; execute Windows containment CI and full timing gate before enabling that platform
W1 original-model IR	Linear numerical model, lifecycle, indicator metadata, six native globals and scoped owning FlatZinc capture/reconstruction implemented; broad capture admission remains	Expand native global registry and capture; measure original-model construction memory
W2 result/session contract	Results, budgets, capability query, persistent HiGHS sessions and owning continuous LP dual/basis observations and owning explicit basis submission implemented; cumulative backend usage remains	Interrupted-workflow accounting and action callback contracts; native persistent state
W3 usable LP/MILP route	First vertical slice implemented and tested; supported I/O dialect is explicit	Broader LP/numerical/general-integer panels, backend-alone comparisons and unsupported-format fixtures
W4 native and FlatZinc integration	Bounded integer linear/indicator native bridge and six typed globals; owning FlatZinc capture with explicit CLI and opt-in MiniZinc registration implemented; broader predicates remain	Broader preserved globals with exact raw-source checks and external CP panels
W5 native sparse relaxation	CSR binary and bounded-integer certificates/interval cuts, native propagation and explicit public hybrid bridge implemented and independently tested	Measure root/whole-solve benefit across families before changing defaults; expand exact arithmetic portability
W6 presolve/postsolve	Exact bounded-integer interval reductions and fixed substitution with owning reconstruction implemented; numerical route also uses HiGHS presolve	Expand mapped constraints and reductions, budgeted solve composition and measured default selection
W7 native branch-and-cut	Proof-carrying scoped pool, binary cover separator, verified root loop and explicit native LP/frontier handoff implemented; dynamic local cuts and broader separation remain	Broader separators, dynamic local-cut lifecycle and measurable root/whole-solve benefit
W8 native branching/global bounds	Explicit frontier/global bounds and opt-in bounded binary reliability probes implemented; broader branching and measured default selection remain	General integer/LP-informed scoring and whole-solve benefit across families with protected regressions
W9 primal portfolio	Numerical starts, exact complete native starts and one explicit bounded binary incumbent neighborhood implemented; partial native completion and repair/diving/RINS/LNS scheduling remain	Cold-start first-feasible gains, broader neighborhood policies and measured benefit with original-model checking
W10 learning/explanations	Architecture/source investigation, isolated pinned SAT protocol experiment and private uncompiled reason-contract header retained; 29,160 experiment queries pass normally and with full dependency sanitizers; no product learning runtime	Checked original-row conditional reason kernel; clone/recompute versus trailed learning pilot; representative benefit before defaults
W11 application workflows	Edits, starts, bounded logic, ordered objectives, numerical grouped conflicts, weighted repair, ranked pools, serial scenarios and selected-basis coefficient/equality sensitivity implemented; fine-grained indicator IIS, broader sensitivity and action callbacks remain	Workflow-by-workflow conformance and a backend/feature support matrix
W12 distribution/Python	C++ packaging, CLI and CI configuration implemented; versioned C ABI/Python including native globals, workflows and LP evidence; local macOS arm64 wheel implemented; other platforms and executed cross-platform CI remain	Expand typed/workflow binding surface, clean platform builds and installed consumers
W13 quadratic/nonlinear	Bounded weighted-square continuous QP with C/Python APIs and independent primal/KKT/bound checking implemented; general Hessians, QCP, MIQP and nonlinear support remain	Broader numerical QP panels and carefully scoped convex quadratic constraints before mixed-integer/nonconvex expansion
W14 native parallel hybrid	Not started	Worker-local state, sound shared bounds/cuts, race/lifecycle testing and uncontended scaling measurements

The seventh integrated checkpoint covers basis submission, Regular and C++ scenario execution, including sanitizer, installed-consumer and wheel checks. Its FAST gate passed 32 cases in 5.69 seconds external wall time. The eighth checkpoint adds LP evidence, scenario C/Python bindings and MiniZinc registration, with full sanitizer, installed-consumer and wheel gates. Its FAST gate passes all 33 cases in 5.73 seconds external wall time, including analytic ray and Farkas checks. Exact source and artifacts are listed in the validation record. The ninth checkpoint integrates selected-basis LP sensitivity with C/Python bindings and bounded native incumbent neighborhoods. It passes the enlarged 35-case FAST gate in 5.7852 seconds, all three CTest configurations, full native sanitizers and fresh installed consumers. Its bounded old/new comparison finds no repeated material case regression among five complete common-workload pairs; one additional unchanged native case hits its existing deadline. Full results and the timing limitation are in the checkpoint report. The separate learning investigation has a checked external-propagation interface experiment; it is not yet a native learning engine. Default performance selection remains outstanding. Broad capture, cold-start primal portfolios, LCG, quadratic constraints/nonlinear support, fine-grained application workflows and native parallel hybrid search remain substantial unfinished work. This ledger does not declare parity from surface API availability.

The full research/source comparison and work-package acceptance criteria remain in the workspace's planning/solver-parity-plan.md, feature-matrix.md, and research-agenda.md. This ledger records implemented scope rather than marking those larger packages complete on the strength of a few successful fixtures.

Parallel work and contention

Model, result/budget and validation work first ran in independent worktrees. The next independent pieces were indicators/Boolean logic, multiobjective coordination and I/O. Each passed its own checks before cherry-picking into this branch. Reviewers then checked one another's work; concrete defects became regression tests before the integrated suite was rerun.

No compilation or solver tests started until the preexisting size-scaling run reported all 3,240 rows complete and its processes exited. Later builds used bounded parallelism. All new outputs live under the isolated implementation build directories; the existing benchmark suite was only read by the new correctness harness. The five-model checks are explicitly not performance measurements or evidence of native/commercial speed parity.