Adaptive native size-boundary method
The adaptive study has sixteen focused generator and runner tests covering preserved input blocks, larger original-model validation, independent version searches, confirmation and budget gates. Each version stops expanding after a failed size and searches its remaining bracket; later failed confirmations remain recorded. Independently checked witnesses establish original feasibility and objective values. Exact optimality remains a backend claim; there is no independent reference-optimum computation during adaptive trials. The earlier fixed studies retain their own measurements and validation records.
Seven-category native scaling measurement
The category scaling study records all 132 observations from 33 frozen cases, two paired repetitions and ten-second solver limits. All returned original-model witnesses validate. Before reports 50 Optimal and 16 TimeLimit results; current reports 58 Optimal and 8 TimeLimit. Whole invocation: 277.4355438 seconds. Knapsack's main size requires both distributions and both repeats, rising from 32 to 384 model variables. Assignment, facility, bin packing, production, routing and queen maxima remain unchanged. The single-variant 512-item stress group is retained separately and remains capped.
Eight new generator tests and nine runner tests pass. Solver source/product and the frozen ten-second driver are unchanged; loader and final integrity checks pass. The earlier 52-run capacity study below remains a separate study, not additional repetitions of this panel. No censored speed ratio or true maximum problem size is claimed.
Native capacity scaling measurement
The capacity scaling study records all 52 observations from 13 frozen cases, two alternating paired repetitions and ten-second internal solver limits. All 52 returned witnesses pass independent original-model checks. Before proves 16/26 optima; current proves 24/26. Largest tested size proved twice rises from 32 to 384 variables in both knapsack distributions, while assignment remains at 256. The 512-variable stress point stays in the capacity calculation; both versions time out in both repetitions. Whole invocation: 150.1307507 seconds.
Eighteen generator/runner tests pass in 1.430 seconds. All five current runtime library hashes match the previous algorithm checkpoint; solver source and product are unchanged. A newly frozen driver accepts the ten-second argument for both cohorts, preserving the older study's driver. The public record includes every observation, frozen model/reference, provenance and integrity result. No timing ratio uses an unfinished solve, and no true maximum size is inferred.
Bounded native capacity algorithm
The algorithm checkpoint supersedes the older current-version wording below. It adds capped exact binary capacity optimization and preserves generic search defaults after rejecting an ordering pilot. Current Release CTest passes 70/70, including Python 84/84; the affected native ASan/UBSan gate passes 12/12, and the new backend-disabled boundary case passes. The independent native test covers 1,494 oracle/budget configurations, with separate deterministic mid-table cancellation/allocation checks. The original checkout still matches all 2,963 recorded hashes. Final FAST passes 35/35 in 5.8439 seconds; capture fixtures pass 13/13. All 60 paired benchmark witnesses pass independent checks: before completes 24/30 runs and after 30/30. The checkpoint retains all timings, control slowdowns and rejected pilot outcomes.
Earlier reports retain their own measured source, configuration and scope.
Final review and PR preparation
The final review corrects public API comments, documents Python's existing four-space convention, and replaces workstation-only documentation links. All C/C++ edits in this follow-up are comments; the measured solver binaries remain unchanged. Python Model.read now allocates its owner before importing a C model, preventing a handle leak when Python allocation fails. The combined library passes all 84 Python tests, including the new allocation regression.
The benchmark reproduction runner now bounds captured output and cleans up process groups on successful exit, timeout, interruption and exceptions. Its seven short Python-fixture tests pass; these tests run no benchmark solver. The website separately pins the audited report hashes before importing its hand-reviewed narrative. Changed evidence is rejected even with Python assertions disabled, without overwriting the public snapshots. These follow-up checks do not replace or alter the measurements below.
Before-project benchmark page and native documentation
The before/after report records a fresh 54-run comparison of preserved pre-project binaries and the ninth integrated product: five binary optimization models and four original native FlatZinc fixtures, three paired repetitions each, two-second solver limits. Known-optimal binary completion is 3/15 before and 15/15 now, with exact checks of returned integer witnesses. The twelve unfinished old runs remain censored. Native CP matches all 12/12 expected-output checks per version, while its tiny end-to-end times are longer in the current shared-library build. No generic speedup is inferred.
The fresh FAST gate passes 35/35 in 5.8003 seconds external wall time. Compiled product remains c2fabb78f; only documentation, benchmark tooling and the separate website change in this batch. The native Doxygen overview, chapters and support matrix were checked against current public APIs. Focused Doxygen 1.18.0 generation passes, and 106 local links/assets/anchors across the new guide and benchmark page resolve. This is not a full original-library documentation build; legacy references outside the focused preview have documented warnings. Both CMake and Autoconf documentation targets include the new source and data.
The website imports recorded measurements without solving; its data checks verify every displayed median, completion count and missing censored time. Production compilation and TypeScript checks pass. A portable export reuses the same page component and built CSS, with the generated native guide and all page data included. Hosting is not claimed while the existing Site is unavailable to the connected account.
Ninth batch: integrated sensitivity and native neighborhoods
The ninth checkpoint report records the completed integration, exact tested source, correctness gates, fresh installed consumers and matched comparison with the eighth checkpoint. Combined CTest passes 67/67, numerical Release 54/54, backend-disabled Debug 52/52, full native/HiGHS ASan+UBSan 66/66, and preloaded sanitizer Python 83/83. Installed consumers pass 8/8 and Python 83/83 in all three configurations. The expanded FAST gate passes 35/35 in 5.7852 seconds external wall time.
Five complete paired runs of the 33 common FAST cases pass 330/330 executions; median command time is 5.6303 seconds previously and 5.7401 seconds currently. No case triggers the repeated >20% and >1 ms diagnostic threshold. One extra attempted current panel times out in unchanged cp_distinct under the existing four-second cap; its failure is retained and excluded from complete-pair timing medians. Three paired five-model MIPLIB panels pass all 30 exact witness and known-optimum checks. See the report for full limitations and artifacts.
This completes the user-requested integration checkpoint. Further roadmap work is on hold. The earlier records below describe their historical checkpoints; in particular the eighth wheel has not been rebuilt for these additions.
Standalone W10 interface experiment after the eighth checkpoint
The separate tools/research target passes 29,160 queries with both the normal pinned CaDiCaL library and the full dependency/fixture instrumented by ASan+UBSan. Each run finds 14,936 SAT and 14,224 UNSAT queries, checks 3,693 external lemmas by finite enumeration, streams 64 delayed/on-demand reasons, and exercises 168 rejected complete models. Ten reason streams occur after intervening notification/decision events. The eight-variable cancellation and resume fixture additionally checks the resumed original witness. The counters are protocol coverage, not a performance comparison or a Gecode learning result.
Evidence is in build/w10/validation.json, cadical-protocol.json, w10-protocol.json, w10-protocol-sanitize.json, and cmake-*-test.log; the aggregate records the final fixture SHA-256 5f1bd71b7646882f137f128a0286d382412d07339355218635dd95e7676570cb. CTest passed 1/1 in each independent build. Both source/header/library provenance records are retained in build/w10-protocol*/dependency-provenance.txt. The library source is the clean tracked CaDiCaL 3.0.1 commit c60730422e758ef1cebe7aeddf2dda31c996bf04; generated untracked build directories remain in that separate dependency. Its upstream example also passed normally. The full sanitizer build uses -fsanitize=address,undefined, frame pointers, assertions and -O1; leak detection was disabled, with halt-on-error enabled for ASan and UBSan. No Gecode product libraries were rebuilt for this experiment.
The first manual fixture run incorrectly asserted that setting quiet must succeed; a dependency compiled with -DQUIET omits that option. The fixture now tolerates that documented compile-mode difference. Upstream configure also prints errors when creating its convenience src/makefile symlink under a path with spaces. Building explicitly inside each generated build directory succeeds; the tracked upstream source was not patched. All dependency configure/build logs are retained in ../deps/CaDiCaL-3.0.1/build-optimize-{spike,sanitize}.
See the learning design for interpretation, uncovered production failure/lifecycle cases and the next checked-reason implementation. This experiment does not advance the product's eighth integrated checkpoint.
Eighth batch: LP evidence bindings and MiniZinc registration
Compiled functional source bda86186d, with documentation-only head 158dc0b66f02929c9be957cdcad651581dd7be3f for the final FAST and wheel gates, tested 2026-09-05 UTC on the same Apple M1 Pro/macOS/AppleClang host and pinned HiGHS 1.15.1. This checkpoint adds numerical LP evidence recovery, its C/Python bindings, scenario C/Python bindings, and the explicit MiniZinc 2.10.1 route. The sensitivity header/private compile bridge is present but its analyzer and bindings are subsequent work. Native neighborhoods and LCG are also excluded.
| Check | Result | Local evidence |
|---|---|---|
| Combined native/numerical and FlatZinc boundary tests | 60/60 pass | build/native-compat/eighth-expanded-ctest.log |
| Numerical Release | 48/48 pass | build/optimize/eighth-ctest.log |
| Backend-disabled Debug | 47/47 pass | build/optimize-core/eighth-expanded-ctest.log |
| Full-source HiGHS/facade ASan+UBSan | 47/47 pass | build/optimize-sanitize/eighth-expanded-ctest.log |
| Full-source native/HiGHS/facade ASan+UBSan | 59/59 pass; preloaded Python 75/75 | build/native-sanitize/eighth-ctest.log, eighth-python-preloaded-sanitizer.log |
| Actual direct FlatZinc CLI | 106/106 normal and sanitizer | build/{native-compat,native-sanitize}/eighth-final-cli.json |
| Actual generated MiniZinc build registration | 72/72 normal and sanitizer, including compiler/driver/output pipeline; pure JSON encoder 10 cases | build/{native-compat,native-sanitize}/eighth-final-minizinc.json, CTest logs |
| Installed MiniZinc registration | 72/72 with actual relative executable/library paths; existing Gecode registration unchanged | build/native-compat/eighth-final-installed-minizinc.json |
| Legacy FlatZinc with blackbox fixtures | 140/140, five repetitions each, seed 1701 | build/native-compat/eighth-legacy-flatzinc.log |
| Installed C/C++ consumers | Numerical, combined and core: 7/7 each; independent native-only 1/1 | build/optimize/eighth-installed-consumers.json |
| Python against installed libraries | 75/75 in each of those three configurations | build/{optimize,native-compat,optimize-core}/eighth-installed-*.log |
| Self-contained macOS arm64 wheel | Automatic Native/HiGHS load and 75/75 on both Python 3.12 and 3.14 | build/eighth-wheel-pack.json, eighth-wheel-installed-results.json |
| Runner and containment contracts | 24 pass, 9 actual-Windows checks explicitly skipped | build/native-compat/eighth-harness-tests.log |
| Required FAST gate | 33/33 pass, 5.73 seconds external wall, 5.68 inside runner | build/optimize/eighth-fast.json, eighth-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 exact witness and known-optimum checks pass | build/optimize/eighth-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, branch and commit unchanged | build/optimize/eighth-original-isolation.json |
The new FAST case checks original feasible bases and improving directions for both objective senses, fixed-column and row recession conditions, historical ownership, private auxiliary identities and public solve-call counts. Its Farkas case independently checks signed multipliers y=1,z=-1, finite-side selection and a positive contradiction margin. The output is evidence_checked, with no source objective values invented from slopes or margins. The full suite retains the 28-second outer allowance and hashes both executables and twelve libraries. All three agents explicitly paused builds/tests and the root test processes exited before timing. A read-only process inventory found no competing task compiler or solver; ordinary desktop background processes remained. This is a correctness timing gate, not a commercial-solver performance comparison.
Evidence conformance covers independently checked original coordinates and separate primal-ray/Farkas availability, bad auxiliary identity/scalars/status, finite-side normalization, tolerance overlap and phase/final cleanup failures. The C wrapper's separately compiled cleanup fixture cancels after a real accepted analysis returns: every accepted getter becomes unavailable while raw owning auxiliary-stage children remain usable. No auxiliary stage becomes an ordinary original-model result. C99 fixtures explicitly preserve assertions in Release.
MiniZinc tests use the official workspace-local 2.10.1 compiler and matching standard library. CMake never downloads it. The generated experimental identity uses Native/Exact with its separate millisecond protocol and dedicated lowering library. Both build and installed registration files are exercised in place; configuration and library hashes must remain unchanged. The existing installed gecode.msc matches the seventh checkpoint byte-for-byte (SHA-256 18e34e3d2dfe96d77e6db410ade28aad2e30c629701c622fb5585e73ab1be463). No default solver preference was changed. Supported source-level answers and emitted predicates are tested; general compiler equivalence is not proved.
The wheel is gecode_optimize-0.1.0.dev20260905+g158dc0b66-py3-none-macosx_11_0_arm64.whl, SHA-256 f1506dc6b1c629e104e8903ee5c9960cdeb6978ed5eb42bb0a5b556d79a377a5. Its solver-library SHA-256 is 67cd47986b72fc91fe3c1f5d8b883b911fda81f2193bbe65d81a00b77e52c006. The first manual Python 3.12 provenance assertion used a nonexistent private loader name and failed before tests ran. Correcting that validation command required no package/source change; final logs assert both module and loaded library are inside the isolated environment, then pass all 75 tests on each interpreter. The failed setup log is retained.
No package-index publication, remote CI, actual Windows/Linux runtime gate, oldest-macOS execution or broad performance baseline is claimed. macOS sanitizer leak detection remains disabled for the host runtime limitation; ASan/UBSan error detection is enabled. LP evidence is numerical, not an exact certificate.
Seventh batch: explicit basis starts, Regular, scenarios and Python wheels
Compiled facade through 77ee30327 and final installed-consumer/FAST source b8e3fba7f152a14c4460f3c6946c488ce59093bc, tested 2026-09-05 UTC on the same Apple M1 Pro/macOS/AppleClang host with pinned HiGHS 1.15.1. The first full cohort used 5914ee047; subsequent fixes keep C consumer assertions enabled, classify the native Regular FAST case correctly, reject unsupported scenario guarantees at admission and accept three exact MiniZinc context-marker atoms. The affected component tests and installed consumers were rerun after those fixes. All facade and test objects were rebuilt for the enlarged Regular GlobalPayload; no old-layout facade archives were reused.
| Check | Result | Local evidence |
|---|---|---|
| Combined native/numerical plus original regression target | 55/55 pass; assertions-enabled C consumers 2/2; final affected tests 7/7 | build/native-compat/seventh-integrated-ctest.log, seventh-c-assertions-ctest.log, seventh-admission-ctest.log |
| Numerical Release | 43/43 pass; final affected tests 4/4 | build/optimize/seventh-integrated-ctest.log, seventh-admission-ctest.log |
| Backend-disabled Debug | 42/42 pass; final affected tests 4/4 | build/optimize-core/seventh-integrated-ctest.log, seventh-admission-ctest.log |
| Full-source HiGHS/facade ASan+UBSan | 42/42 pass; final affected tests 4/4 | build/optimize-sanitize/seventh-integrated-ctest.log, seventh-admission-ctest.log |
| Full-source native/HiGHS/facade ASan+UBSan | 52/52 pass; final affected tests 7/7; preloaded Python 62/62 | build/native-sanitize/seventh-integrated-ctest.log, seventh-admission-ctest.log, seventh-python-preloaded-sanitizer.log |
| Actual FlatZinc CLI | 106/106 in normal and full sanitizer configurations | build/{native-compat,native-sanitize}/seventh-final-cli.json |
| Legacy FlatZinc with configured blackbox fixtures | 140/140 pass | build/native-compat/seventh-final-legacy-flatzinc.log |
| Installed C/C++ consumers | Three configurations: 5/5 each, including basis/Regular C consumers and scenario C++ behavior; independent native-only 1/1 | build/optimize/seventh-installed-consumers.json, seventh-final-installed-results.json |
| Python against installed libraries | 62/62 in each of three configurations | build/{optimize,native-compat,optimize-core}/seventh-final-installed-*.log |
| Self-contained macOS arm64 wheel | Automatic Native/HiGHS load, RECORD/bytecode checks and 62/62 methods on both Python 3.12 and 3.14 | build/seventh-wheel-integrated-pack.json, seventh-wheel-integrated-installed-results.json |
| Runner and containment contracts | 24 pass, 9 actual-Windows checks explicitly skipped | build/native-compat/seventh-harness-tests.log |
| Required FAST gate | 32/32 pass, 5.69 seconds external wall, 5.62 inside runner | build/optimize/seventh-fast.json, seventh-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 independent exact witness/known-optimum checks pass | build/optimize/seventh-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, branch and commit unchanged | build/optimize/seventh-original-isolation.json |
The FAST run had a clean source head, explicit quiet acknowledgments from every agent and a process inventory with no competing compiler or solver. It hashes both executables and twelve runtime libraries. New cases independently enumerate a signed Regular parity language (exact objectives -3/1) and check analytic scenario objectives 5/9/-11/7 with both cold and reused solves. The LP observation case now also exports and submits an owning basis. The outer budget stays at 28 seconds. This is a bounded correctness regression gate, not evidence of commercial or native performance parity.
Regular admission/checking covers sparse large state namespaces, signed symbols, missing edges, aliases, empty words/finals, determinism, native array/hash limits, FlatZinc literal schemas and original raw-table checks. Scenario tests compare 360 outcomes with 3,840 independent original assignments, and inject bad identity, status, exact bounds, raw witnesses and interruptions. Unsupported Certified and Auto/HiGHS Exact requests are rejected before a batch starts. Basis tests cover status/owner/revision/matrix compatibility, rejected or repaired submissions, original optimality checks and session recovery.
The initial new C consumer targets inherited Release NDEBUG; their source now explicitly enables assertions and their Release checks were rerun. The first manual legacy selection omitted optional blackbox environment variables and selected only 118 tests. A second attempt used an incorrect local fixture-library path and stopped at that fixture; the final run uses the configured executable and library paths, separate log files and seed 1701, passing all 140 tests. These setup attempts are retained in the local logs.
The wheel artifact is gecode_optimize-0.1.0.dev20260905+g2e19db01d-py3-none-macosx_11_0_arm64.whl (SHA-256 108a5b81b57aa6c6ebe66c1b668fa4bd3f0d3ffe6b418aa62eba72dae1a16701). It was built with macOS 11.0 deployment settings and tested on this newer host; oldest-OS compatibility is not claimed as tested. Ten portable adversarial packager tests also pass, including staged-byte inspection during source replacement, dependency traversal and corrupt/missing RECORD entries. No package index publication, actual Windows execution, remote CI or broad performance baseline was performed. macOS sanitizer leak detection remains disabled because of the host runtime limitation; ASan/UBSan error detection remains enabled.
Scenario C/Python bindings, explicit LP evidence and experimental MiniZinc registration are subsequent isolated work and are excluded from this checkpoint.
Sixth batch: starts, original LP observations and preserved FlatZinc globals
Integrated source through 5e22c417f6824981f8a705b29af1b7c352717f10, tested 2026-09-05 UTC on the same Apple M1 Pro/macOS/AppleClang host with pinned HiGHS 1.15.1. Compiled changes are through 7364d7a77; subsequent commits add cumulative CLI fixtures and scenario design only. This gate adds exact complete native starts, owning continuous LP observations with C/Python access, and FlatZinc table/holey-domain, circuit and fixed-data cumulative admission. Basis submission, scenario execution and regular-language constraints are separate subsequent work and are excluded from this gate.
| Check | Result | Local evidence |
|---|---|---|
| Combined native/numerical plus original regression target | 48/48 pass | build/native-compat/sixth-observations-ctest.log |
| Numerical Release | 36/36 pass | build/optimize/sixth-observations-ctest.log |
| Backend-disabled Debug | 36/36 pass | build/optimize-core/sixth-observations-ctest.log |
| Full-source HiGHS/facade ASan+UBSan | 35/35 C/C++ tests pass | build/optimize-sanitize/sixth-observations-ctest.log |
| Full-source native/HiGHS/facade ASan+UBSan | 45/45 optimization/capture tests pass; 47 Python methods pass with runtime preloaded | build/native-sanitize/sixth-observations-final-ctest.log, sixth-python-preloaded-sanitizer.log |
| Expanded actual FlatZinc CLI | 79/79 cases pass in normal and full-source sanitizer builds | build/{native-compat,native-sanitize}/sixth-cumulative-cli-final.json |
| Complete legacy FlatZinc selection | 140/140 pass | build/native-compat/sixth-final-legacy-flatzinc.log |
| Fresh installed C/C++ consumers | Standalone numerical, combined and backend-disabled: 3/3 each; independent native-only: 1/1 | build/consumer-*-sixth/, build/optimize/sixth-installed-consumers.json |
| Python against freshly installed libraries | 47/47 methods pass in each of three configurations | build/{optimize,native-compat,optimize-core}/sixth-installed-*.log |
| Runner and containment contracts | 24 pass; 9 actual-Windows checks explicitly skipped | build/native-compat/sixth-harness-final-tests.log |
| Required FAST gate | 30/30 pass, 5.48 seconds external wall, 5.40 seconds inside runner | build/optimize/sixth-fast.json, sixth-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 independent exact witness/known-optimum checks pass | build/optimize/sixth-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, original branch and commit unchanged | build/optimize/sixth-final-original-isolation.json |
The final FAST run used a clean source head, explicit quiet acknowledgments from all three agents, and a process inventory showing no competing build or solver workload. Both executables and twelve runtime libraries are hashed. The new native-start case checks gate completion and strict min/max improvement to enumerated optima -4/7. The LP observation case checks analytic optima 15/19, original dual signs, reduced costs, slacks, basis availability and independent KKT acceptance. The outer budget remains 28 seconds. This local correctness timing is not a commercial/native performance comparison.
Normal and core/numerical tests were run after source integration. The later CLI-only commit expands 64 cases to 79 without changing the compiled solver; the final 79-case checks were rerun in both normal and sanitizer configurations. Cumulative compiler tests independently evaluate 248,201 assignments. Actual legacy CLI execution rejects a zero-duration singleton with height above capacity; the new path accepts it because its half-open resource interval is empty. The positive-duration counterpart is infeasible. Raw original-source checking and independent integer-time oracles cover both behaviors, aliases, overlap, shared endpoints, malformed shapes and unsupported machine variants.
Native starts are checked before incumbent publication, including exact original feasibility, live indicator gate derivation, explicit retained gates, historical ownership, interrupted search and strict improvement completion. LP observations check original row/column mapping, min/max conventions, large offset cancellation, malformed backend data, independent KKT residuals, basis availability, copied ownership and final-budget publication. Fresh installs include the public LP observation header and exclude its private detail header.
The initial native sanitizer CTest command also selected an unnecessary rebuild of the legacy test harness. That owned process tree was stopped before running the established optimization/capture selection, which passed all 45 tests. Legacy regression evidence in this batch comes from the normal combined build; the solver modules and dependencies used by the sanitizer selection are fully instrumented. No sanitizer detector was disabled beyond the existing macOS leak-detection limitation. The combined installed consumer initially omitted its separately installed HiGHS dependency prefix; configuring it with the fresh standalone-sixth HiGHS install passed. Logs retain both attempts. No remote CI, actual-Windows gate, or broad performance baseline was executed.
Fifth batch: quadratic models, FlatZinc and native search extensions
Integrated implementation 8ae59d5c1946abf47203172ba3cb89fbf5ea0a8b, tested 2026-09-05 UTC on the same Apple M1 Pro/macOS/AppleClang host and pinned HiGHS 1.15.1. Subsequent commits through a747330805cf1fed4c082a42900d9194c9f849b8 change documentation only; that clean head is recorded by the final FAST run. This batch adds verified root-cover handoff, opt-in bounded binary reliability branching, bounded weighted-square continuous QP with C/Python APIs, owning FlatZinc capture, scoped integer/Boolean compilation and the separate fzn-gecode-optimize command. W9 starts, LP observations and broader FlatZinc global admission are subsequent work, excluded from this gate.
| Check | Result | Local evidence |
|---|---|---|
| Combined native/numerical plus original regression target | 43/43 pass | build/native-compat/fifth-final-ctest.log |
| Numerical Release | 32/32 pass | build/optimize/fifth-final-ctest.log |
| Backend-disabled Debug | 32/32 pass | build/optimize-core/fifth-final-ctest.log |
| Full-source HiGHS/facade ASan+UBSan | 31/31 C/C++ tests pass | build/optimize-sanitize/fifth-final-ctest.log |
| Full-source native/HiGHS/facade ASan+UBSan | 40/40 tests pass, including FlatZinc capture/CLI; 37 Python methods pass with runtime preloaded | build/native-sanitize/fifth-final-ctest.log, fifth-python-preloaded-sanitizer.log |
| Complete legacy FlatZinc selection | 140/140 pass | build/native-compat/fifth-final-legacy-flatzinc.log |
| Fresh installed C/C++ consumers | Standalone numerical, combined, backend-disabled: 2/2 each; independent native-only: 1/1 | build/consumer-*-fifth/, build/optimize/fifth-installed-consumers.json |
| Python against freshly installed libraries | 37/37 methods pass in each of three configurations | build/{optimize,native-compat,optimize-core}/fifth-installed-python.log |
| Runner and containment contracts | 24 pass; 9 actual-Windows checks explicitly skipped | build/native-compat/fifth-harness-final-tests.log |
| Required FAST gate | 28/28 pass, 5.55 seconds external wall, 5.52 seconds inside runner | build/optimize/fifth-fast.json, fifth-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 independent exact witness/known-optimum checks pass | build/optimize/fifth-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, original branch and commit unchanged | build/optimize/fifth-final-original-isolation.json |
FAST ran after every root build/test process exited and all three agents confirmed a quiet window. A process inventory found no build/solver workloads; both executables and twelve runtime libraries are hashed. The unchanged outer budget is 28 seconds, including startup, hashing, checking and cleanup. The 5.55-second observation is a local correctness gate, not a solver-performance comparison. The new cases independently check QP optima 3.5/2.5, root-cover optima 15/19 and reliability-branching optima -17/17, including actual cut/probe accounting rather than merely accepting an optimal status.
The expanded native sanitizer build initially exhausted disk space while linking FlatZinc executables. Superseded generated agent objects/archives and executables were removed; source, current inputs and validation logs were preserved. The serialized retry built all targets and the final 40-test gate passed. Native, integer/set/float, minimodel, search, FlatZinc, HiGHS and facade sources were instrumented with matching generated configuration. The original checkout was not used as a build/output directory. macOS leak detection remains unavailable (detect_leaks=0); address and undefined-behavior checks halt on errors. No additional detector was disabled.
Broader legacy testing exposed one existing restart-output expectation: on_restart_sol_float required y=1.0 for an initially unconstrained finite float. Both the frozen fourth-batch executable and the current one printed -DBL_MAX there. The test now validates any finite first value and still requires the exact later remembered values, sequence and completion markers; adversarial output fixtures check that it rejects malformed or semantically wrong output. The model and search settings are unchanged. This test-only correction is included in the 140/140 final legacy result; it is not evidence of a solver behavior regression or a relaxation of later restart checks.
Quadratic tests validate original primal values, objective/gradient, dual signs, KKT residuals and an independently checked outward tangent/box bound, including large-offset cancellation, interrupted backend data and fast-math rejection. They do not claim arbitrary-Hessian, mixed-integer quadratic or nonlinear support. Native reliability tests use exhaustive finite-product oracles and fault/interruption checks; probes rank candidates but never publish proofs or incumbents. FlatZinc tests use independent raw-source predicates and whole-model admission, including CLI status/identity/output corruption fixtures. Fresh installs contain the public quadratic header and exclude its private bound header. No remote CI or actual-Windows containment gate has run locally.
Fourth batch: checked native search, presolve, pools and bulk bindings
Integrated source cf95a6167342953ddb386a9a87b18e58c7672eb5, tested 2026-09-05 UTC on the same Apple M1 Pro/macOS/AppleClang host and pinned HiGHS. This gate includes native LP deductions, scoped cover proof checking, explicit native search with interrupted frontier bounds, exact integer presolve/postsolve, ranked pools, C/Python pool and repair workflows, and atomic C++/C/Python batches. The root cut loop and subsequent QP/FlatZinc work are separate later slices.
| Check | Result | Local evidence |
|---|---|---|
| Combined native/numerical plus original regression target | 31/31 pass | build/native-compat/fourth-final-ctest.log |
| Numerical Release | 26/26 pass | build/optimize/fourth-final-ctest.log |
| Backend-disabled Debug | 26/26 pass | build/optimize-core/fourth-final-ctest.log |
| Full-source HiGHS/facade ASan+UBSan | 25/25 C/C++ tests pass | build/optimize-sanitize/fourth-final-ctest.log |
| Full-source native/HiGHS/facade ASan+UBSan | All 28 C/C++ tests pass; 28 Python methods pass with the sanitizer runtime preloaded | build/native-sanitize/fourth-final-ctest.log, fourth-python-preloaded-sanitizer.log |
| Fresh installed C/C++ consumers | Standalone numerical, combined, backend-disabled: 2/2 each; independent native-only: 1/1 | build/consumer-*-fourth/, build/optimize/fourth-installed-consumers.json |
| Python against freshly installed libraries | 28/28 methods pass in each of standalone numerical, combined and backend-disabled configurations | build/{optimize,native-compat,optimize-core}/fourth-installed-python.log |
| Runner and containment contracts | 24 pass; 9 actual-Windows checks explicitly skipped | build/native-compat/fourth-harness-final-tests.log |
| Required FAST gate | 25/25 pass, 5.57 seconds external wall, 5.50 seconds inside runner | build/optimize/fourth-fast.json, fourth-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 independent exact witness/optimum checks pass | build/optimize/fourth-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, original branch and commit unchanged | build/optimize/fourth-final-original-isolation.json |
All root and agent builds/tests had exited before FAST ran, with explicit agent acknowledgments and a saved process inventory. Both executables and twelve runtime libraries are hashed. Normal desktop services remained; this single correctness-gate observation is not a solver-performance comparison. The outer runner budget remains 28 seconds. No original benchmark output was modified.
The native sanitizer CTest invocation initially also registered an ordinary Python host. That one launch failed before the tests because AddressSanitizer loaded too late through ctypes. The native C/C++ tests all passed. Running the Framework Python application executable with the matching sanitizer runtime preloaded then passed every Python test; the ordinary Python registration is disabled in this sanitizer build. No sanitizer detector was suppressed. Local leak detection remains unavailable (detect_leaks=0); address, container and undefined-behavior checking use halt-on-error.
Native LP tests independently enumerate 320 configurations. Frontier tests exercise complete and interrupted minimization/maximization, queued and active parent-region coverage, storage and node limits, cancellation/deadlines, allocation failures, original witness checks, and optional checked LP. The separate coordinator executable compiles test hooks into its own implementation; the production library contains no test hooks. Presolve tests enumerate full original/reduced feasible sets and add actual native/HiGHS reconstruction checks. Bulk tests inject allocation failures before every allocation in all three C++ operations, verify unchanged revisions/contents/reference addresses, and check sparse construction without dense allocation. C/Python tests cover output capacity, foreign/deleted handles, UTF-8, sparse dimensions and owning historical workflow results. These are correctness gates; default native performance selection and broader benchmark-family comparisons remain outstanding.
Third batch: globals, sparse LP, repairs, bindings and native lifecycle
Integrated source through 0c6d9462d, tested 2026-09-05 UTC on Apple M1 Pro, macOS 15.7.4 and AppleClang 17. The HiGHS pin remains unchanged. Later work, including solution pools and bounded-integer sparse certificates, has separate acceptance checks and is not included in this recorded gate.
| Check | Result | Local evidence |
|---|---|---|
| Numerical Release conformance | 19/19 pass, including 13 Python conformance methods | build/optimize/third-batch-ctest.log |
| Backend-free conformance | 19/19 pass; expanded binding checks also pass after their final addition | build/optimize-core/third-batch-final-ctest.log, third-batch-bindings-ctest.log |
| Combined native/numerical and original regression target | 22/22 pass | build/native-compat/third-batch-final-green-ctest.log |
| Full HiGHS and facade ASan/UBSan | 18/18 pass with halt-on-error; Python host excluded from this instrumented CTest configuration | build/optimize-sanitize/third-batch-final-ctest.log |
| Full native engine ASan/UBSan | 4/4 pass: globals, native solve, brancher lifecycle, C ABI; 8.22 seconds | build/native-sanitize/third-batch-final-ctest.log |
| Clean installed C/C++ consumers | Standalone numerical, combined native/numerical, backend-free and separate native-only consumers pass | build/consumer-*-third/, local /private/tmp/gecode-third-installed-*.log |
| Installed Python | 13/13 methods pass against standalone numerical and combined native/numerical libraries | build/optimize/third-batch-installed-python.log, build/native-compat/third-batch-installed-python.log |
| Package ownership | Six collision/repeat cases plus standalone-native boundary pass | build/native-compat/third-batch-package-origin.log |
| Fast runner contracts | 11 tests pass, including new native-global provenance checks | build/native-compat/third-batch-final-runner-tests.log |
| Required fast gate | 21/21 pass; 5.53 seconds external wall, 5.45 seconds inside the runner | build/optimize/third-batch-fast.json, third-batch-fast-wall.json |
| Existing MIPLIB correctness panel | 5/5 pass with independent original witness checker; clean recorded source tree | build/optimize/third-batch-miplib-correctness.json |
| Original checkout isolation | All 2,963 input/source hashes plus original branch and commit unchanged | build/optimize/third-batch-original-isolation.json |
The fast command ran after build/test processes finished and agents confirmed a quiet window. It includes all startup, validation, cleanup and output work, both executable hashes and twelve shared-library hashes. This is one local correctness-gate timing, not a solver-performance comparison. All required cases must finish within the unchanged 28-second runner budget. See the fast panel for independent schedule, recourse, repair and native integer oracles.
Native global tests enumerate finite products with independent predicates, including aliases, negative index bases, empty tables/elements, zero-resource tasks and circuit subtours. Resource preflight now bounds native envelope, edge-finding and index arithmetic; targeted extreme-value fixtures must reject unsupported ranges before calling those propagators. The sparse binary LP slice has independent 3,000-model dense/CSR certificate comparisons, 8,640 native propagation configurations and a 16K-square allocation guard; see sparse LP for exact scope and the isolated-worktree instrumentation used for those tests.
Cross-review fixed residual cancellation across large row bounds and objective offsets, invalid stage evidence, and lost raw-stage provenance on interruptions. Repair tests use 21 deterministic coordinator scenarios in addition to actual HiGHS and exhaustive small-model checks. Repairs retain a distinct private model and cannot claim that an original infeasible point became feasible.
Full native instrumentation first exposed invalid downcasts of intrusive list sentinels to Brancher/Propagator. The kernel now retains ActorLink cursors and casts only actual actors after the sentinel check. Lifecycle tests cover empty lists, forward and wrapped lookup, absent middle actors, deletion, archive choices, cloning, propagator iteration and recomputed DFS solution sets. Final runs retain the vptr detector and use UBSAN_OPTIONS=halt_on_error=1; no undefined-behavior detector was suppressed. macOS leak detection is unavailable, so ASAN_OPTIONS=detect_leaks=0:halt_on_error=1 was used locally. The added Linux native sanitizer job enables leak detection, but remote CI has not been run here.
All native components and consumers must be rebuilt together after the cursor representation change. The C layout remains independent of C++ internals. ELF static native dependencies are compiled as PIC when embedded in the shared C wrapper. A native-enabled binding fixture incorrectly expected a numerical session to fall back to native execution; it now checks the documented Unsupported result. Standalone native/global calls pass through both bindings.
Second batch: sessions, native bridge, diagnostics and fast gate
Tested solver implementation: 6a7544e85, on codex/solver-parity, 2026-09-05 UTC. Host: Apple M1 Pro, macOS 15.7.4, AppleClang 17. HiGHS remains pinned to 1.15.1, commit 04024d701f79feb8e2f18bc3df0dffc04ef05088.
| Check | Result | Local evidence |
|---|---|---|
| Numerical Release conformance | 14/14 pass | build/optimize/integrated-ctest.log |
| Backend-free conformance | 14/14 pass | build/optimize-core/integrated-ctest.log |
| Combined native Gecode and optimization | 16/16 pass, including the enabled native bridge and original native regression target | build/native-compat/integrated-ctest.log |
| Full-source Debug ASan/UBSan, including HiGHS | 14/14 pass; 41.61 seconds for the complete conformance run | build/optimize-sanitize/integrated-ctest.log |
| Four installed consumers | Standalone numerical, combined native/numerical, native-only without HiGHS discovery, and backend-free all pass | build/consumer-*/integrated-ctest.log |
| Fast regression command | 18/18 required cases pass; 5.49 seconds external wall time, 5.37 seconds reported inside runner, against a 28-second budget | build/optimize/second-batch-fast.json, second-batch-fast-wall.json |
| Fast-runner failure regressions | 11 tests pass | build/optimize/fast-runner-tests.log |
| Package ownership/support boundaries | Four configure-only ownership cases plus standalone-native rejection pass | build/optimize/package-origin-tests.log |
| Existing MIPLIB correctness inputs | 5/5 pass with the original separate exact witness checker | build/optimize/second-batch-miplib-correctness.json |
| Original checkout isolation | All 2,963 hashes, original branch and commit unchanged | build/optimize/second-batch-original-isolation.json |
The fast measurement ran alone after builds and solver tests finished; no original benchmark process was observed. It includes Python startup, child startup, source/binary hashing, eleven shared-library hashes, all cases, validation, cleanup and report writing. This is one local timing observation, not a performance ranking or a guarantee about slower machines. Every required case must complete within the runner's outer budget or the command fails. The native bridge case has explicit native backend/version and Exact guarantee attribution; its two optima are independently enumerated.
The new tests compare persistent edits with cold solves and exhaustive integer oracles, exercise active/zero semi domains and indicators, and check bounds, costs, objective senses, matrix/type/owner changes, starts and historical values. A nontrivial 48-binary MIP is still unresolved after one node both cold and with a retained hint; removing the node-limit reset in a temporary mutation makes the unlimited-session regression fail. Node-limit termination after callback registration also exercises callback teardown before a fresh model. That conformance test has a separate 180-second timeout for slow Debug sanitizer builds and is not part of the fast gate.
Conflict tests reconstruct reported subsets independently and check deletion witnesses; deterministic oracle substitutes cover ambiguous and interrupted results. Native bridge tests exhaustively check finite integer models, reification, offsets, semi domains, unsupported ranges and interrupted search. Separate native-bridge ASan/UBSan tests instrumented the bridge and its test sources; the legacy native libraries used by those tests were not instrumented. The full-source HiGHS sanitizer configuration uses the disabled native bridge stub. A full native-library sanitizer configuration remains additional coverage.
Integration found a Debug HiGHS assertion when a diagnostic reduction left integer variables and only constant rows. The adapter now decides constant rows using the original numerical checker's absolute feasibility tolerance and omits them from the backend matrix, while retaining them in the original model. The regression covers feasible, contradictory and tolerance-boundary constants. The full-source sanitizer run passes without suppressing backend assertions.
The package tests prevent silent target collisions between different standalone prefixes or a combined and standalone installation. The added Linux static and Windows shared native CI configurations were checked locally at configure time; their remote CI jobs have not run. Windows fast-runner process containment is still pending. Generated build reports are local artifacts, not committed benchmark results. The MIPLIB report's dirty flag reflects temporary Python bytecode files from the runner tests, subsequently removed; solver source was committed at the recorded hash.
First-batch evidence
Completed on 2026-09-05 UTC (2026-09-04 local time). Tested implementation commit: 8c45e7b30, on codex/solver-parity. Documentation was completed afterward. Host: Apple M1 Pro, macOS, AppleClang 17, C++17. Backend: separately pinned HiGHS 1.15.1, commit 04024d701f79feb8e2f18bc3df0dffc04ef05088.
First-batch checks
| Configuration or check | Result | Local evidence |
|---|---|---|
| Standalone numerical Release build | 9/9 CTest targets pass | build/optimize/final-ctest.log |
| Full-source Debug ASan + UBSan, including HiGHS | 9/9 pass, no sanitizer suppression | build/optimize-sanitize/final-ctest.log |
| Standalone backend-disabled build | 9/9 pass | build/optimize-core/final-ctest.log |
| Combined native Gecode + optimization | 11/11 pass, including native regression target and its build fixture | build/native-compat/final-ctest.log |
| Installed standalone numerical consumer | Configure, build and runtime pass | build/consumer-standalone/final-ctest.log |
| Installed native + numerical consumer | Configure, build and runtime pass | build/consumer-combined/final-ctest.log |
| Installed native-only consumer, HiGHS discovery disabled | Configure, build and runtime pass | build/consumer-native-only/final-ctest.log |
| Installed backend-disabled optimization consumer | Configure, build and expected Unsupported runtime contract pass | build/consumer-core/final-ctest.log |
| Explicit native static-only selection | gecodeoptimize builds as a static library, honoring Gecode linkage options | build/static-configuration/build.log |
| Existing five-model MIPLIB correctness harness | 5/5 pass, independently exact-checked witnesses and known optima | build/optimize/final-miplib-correctness.json |
| Original checkout isolation | All 2,963 file hashes, original branch and original commit unchanged | build/optimize/original-isolation.json |
| Changelog tidy and whitespace | Changelog checks pass; git diff --check passes | Checker details below |
| Added platform CI | YAML and all shell steps checked locally; remote jobs not executed | .github/workflows/optimize.yml |
The build logs and JSON reports are generated local artifacts, not committed benchmark results. The MIPLIB report records the binary hash, input hashes, backend version and tested source commit. Its dirty-worktree flag reflects the implementation-ledger documentation being edited during the run; solver code was committed at the hash above.
The nine component targets cover model lifecycle; result/gap/budget semantics; independent validation; indicators/Boolean logic; I/O; actual solver integration; ordered objectives; public-file I/O; and a deterministic multiobjective coordinator with a substitute backend. Assertions remain enabled in Release test executables.
Independent worktree checks preceded integration: normal C++17 builds with -Wall -Wextra -pedantic, then address/undefined sanitizer runs for the owned components. Cross-review fixes were tested again in the combined suite. The actual solver test combines indicators, partial starts, lexicographic locks and indicator removal, in addition to LP/MILP objectives, edits, bounds, limits, semi-variable starts, deleted handles and tiny differential integer models.
An initial sanitizer configuration mixed instrumented adapter code with an uninstrumented static HiGHS library and reported a libc++ vector-container annotation error. Rebuilding both with the same instrumentation passes all tests. Container checking was not disabled. The CI sanitizer job therefore builds the entire dependency from source with matching C and C++ flags.
Existing benchmark inputs
The five unchanged cached MIPLIB3 models were solved cold, with one worker, seed zero, zero native MIP gaps and a 30-second per-case limit. Reference objectives were used for checking only, never supplied as starts or targets.
| Model | Variables | Original rows | Published and independently checked optimum |
|---|---|---|---|
| p0033 | 33 | 16 | 3089 |
| p0201 | 201 | 133 | 7615 |
| p0282 | 282 | 241 | 258411 |
| p0548 | 548 | 176 | 8691 |
| lseu | 89 | 28 | 1120 |
experiments/optimize/verify_miplib.py reads the original MPS using the existing suite's separate exact parser, verifies pinned payload hashes, maps returned values by column name, checks numerical integrality, then checks rounded binary witnesses with the existing exact assignment validator. Independently recovered objectives must match both the published optimum and the adapter's result; reported bounds must close at that optimum. Unsupported, failed or timed-out cases fail the correctness check rather than disappearing from the denominator. The report records external process wall time as well as the CLI's internal timing fields, and rejects nonfinite JSON numbers.
These are correctness checks on five small historical instances. They are not a commercial-solver comparison, a full MIPLIB2017 result, a performance baseline, or evidence for changing native search defaults. The original 3,240-row size-scaling run completed before any implementation compilation or solver testing started. Its suite and output files were not changed by this work.
Import/export regressions
I/O tests compare every active domain, row, coefficient, objective and display name after local round trips, including all five public models through both formats. Backend-enabled tests also read exported files with the independent pinned HiGHS reader and compare their mathematical content; dropping an unconstrained free row is permitted in that external comparison only.
Cross-review reproduced and fixed:
- 15-digit output merging distinct coefficients and changing integer feasibility.
- Lost semi-integer declarations and omitted cost-free columns.
- Duplicate objective terms and LP constants on constraint left sides.
- Cancellation across a large LHS constant and RHS, preserving x >= -1 in x + 1 - 1e16 >= -1e16.
- Conflicting MPS domain declarations silently dropping integrality.
- MPS binary bounds that passed local round trips but were ignored by another reader. Nondefault binary bounds now require LP export; MPS rejection leaves an existing destination unchanged.
- Original-indicator metadata being lost through linear file exchange. Active original indicators are rejected by these exporters.
The independent reviewer reran the cancellation, conflicting-type, destination preservation and narrowed-binary LP fallback reproductions successfully after the fixes. See I/O scope for syntax and portability restrictions.
Reproduction and remaining release work
Build instructions are in the API guide. For a fully instrumented local run, use a fresh build directory and supply the pinned source path:
The changelog-only misc/tidy.py checks were run through installed Python 3.14, with its two uv run generator launches directed to that interpreter because uv was unavailable. The generators are standard-library-only, require Python 3.11+, and their checking logic and source files were unchanged.
Windows and Linux CI execution, C ABI/Python packaging, broad numerical and general-integer panels, and an uncontended cold-start performance baseline remain release work. The added CI workflow covers static/shared core packages on three platforms, numerical packages, installed consumers, and full-source sanitizers, but no remote success is claimed. Native/FlatZinc compilation, native sparse LP/cuts/branching, explanations/learning, quadratic/nonlinear models and native parallel hybrid solving remain separate roadmap work.
Generated by
1.18.0