The work recorded at this historical checkpoint is integrated on codex/solver-parity in implementation/gecode. This checkpoint finishes selected-basis LP sensitivity and bounded native incumbent improvement, validates their integration, and compares existing workloads with the preceding working checkpoint. Further roadmap implementation is on hold. The later bounded algorithm work is recorded separately in the algorithm checkpoint.
Compiled product source is c2fabb78f1fc3bcfd7432eb71cfdb22656cd3f60. The final FAST runner and performance measurement use db7eb0b85b2970e20e8a3d965559fd31a94c73b4; that follow-up changes only the runner's expected BinaryHamming backend identity. Later checkpoint documentation does not change the tested product. The original checkout remains on experiment/llm-search at 3e0e8ee76fb4ba01c53616dce02c0bb1a53f159e, with all 2,963 recorded source/test/input hashes unchanged.
Completed scope
- LP sensitivity computes original-coordinate objective coefficient and equality-RHS intervals for a selected optimal continuous LP basis. The numerical analyzer checks the basis and source observations, factors a private system without running another optimization, and retains owning history and explicit incomplete/unavailable outcomes. C and Python bindings include lifetime, cancellation and cleanup behavior.
- Native neighborhoods adds an explicit C++ API for one bounded binary Hamming-distance improvement attempt after a checked incumbent. It checks a proposed improvement against the original model and shares the outer search budget without promoting a local proof to a global one. Existing default solve routes remain unchanged. This experimental slice does not include a C/Python neighborhood wrapper or a full heuristic portfolio.
- Both features are integrated into normal/backend-disabled/sanitizer builds, installed consumers, and the expanded 35-case FAST gate. Sensitivity's private factorization headers are excluded from installed headers.
- Previously started learning research is retained as an isolated experiment and private reason-contract header. The header is excluded from installation; no CaDiCaL dependency or learning runtime was added to Gecode.
Correctness and installation
All results below are from the local Apple M1 Pro, macOS 15.7.4, AppleClang host, with pinned HiGHS 1.15.1 and MiniZinc 2.10.1 where applicable. These are different configurations of overlapping tests; their counts should not be added as unique test coverage.
| Gate | Result | Evidence relative to repository root |
|---|---|---|
| Combined native/numerical and FlatZinc boundary CTest selection | 67/67 pass | build/native-compat/ninth-ctest.log |
| Numerical Release | 54/54 pass | build/optimize/ninth-ctest.log |
| Backend-disabled Debug | 52/52 pass | build/optimize-core/ninth-final-ctest.log |
| Full native/HiGHS/facade ASan+UBSan configuration | 66/66 pass | build/native-sanitize/ninth-ctest.log |
| Python with the sanitizer runtime preloaded | 83/83 pass | build/native-sanitize/ninth-python-preloaded-sanitizer.log |
| Installed C/C++ consumers | 8/8 each: numerical, core, combined; native-only 1/1 | build/optimize/ninth-installed-consumers.json; build/native-compat/ninth-native-only-test.log |
| Python against installed libraries | 83/83 in each of those three configurations | build/optimize/ninth-installed-consumers.json |
| Actual FlatZinc CLI | 106/106 in normal and sanitizer configurations | build/{native-compat,native-sanitize}/ninth-cli.json |
| Actual MiniZinc build registration | 72/72 in normal and sanitizer configurations | build/{native-compat,native-sanitize}/ninth-minizinc.json |
| Actual installed MiniZinc registration | 72/72 pass | build/native-compat/ninth-installed-minizinc.log |
| Legacy FlatZinc including optional blackbox fixtures | 140 fixture instances, five repetitions each, seed 1701 | build/native-compat/ninth-legacy-summary.json |
| Regression-runner/containment unit tests | 24 pass; 9 actual-Windows tests skipped | build/native-compat/ninth-harness-tests.log |
| Expanded FAST gate | 35/35 pass; 5.7852 s external wall within its unchanged 28 s outer budget | build/optimize/ninth-fast.json, ninth-fast-wall.json |
| Original checkout isolation | 2,963 hashes plus branch and HEAD unchanged | build/optimize/ninth-original-isolation.json |
The new FAST sensitivity case checks 63 analytic/lifecycle conditions for both objective senses. The neighborhood case checks 14 conditions, including an actual incumbent improvement and the independent eight-assignment optimum of −29. Private coordinator tests cover interruptions, accounting and cleanup; installed consumers exercise the public APIs. The original full native test corpus was not rerun: the retained CP gate and the selected legacy FlatZinc fixtures provide the bounded original-code regression coverage.
The first legacy selection omitted the log environment variable, selecting 136 fixture instances. A follow-up with the proper blackbox environment ran the four missing fixtures and repeated seven shared controls. Two distinct legacy fixtures both register as tenpenki::1, so there are 139 distinct display names among the 140 fixture instances. This matches the preceding checkpoint's fixture set. An initial new FAST fixture compile used the wrong LP observation options type; that test setup was corrected before the final rebuild. The runner's expected neighborhood backend identity was also corrected before the final FAST run. Those preliminary failures are retained in local logs.
Existing-workload performance comparison
The prior checkpoint is compiled source bda86186d, with source/report head 158dc0b66f02929c9be957cdcad651581dd7be3f. Both cohorts use the same preserved eighth-checkpoint test executables and the same frozen 33-case runner and 70 fingerprinted source/fixture files. Only explicitly selected solver libraries change. Dynamic-loader probes verified all 11 loaded Gecode libraries came from the intended cohort, and artifact hashes were checked again afterward. All task builds/tests stopped before timing; ordinary desktop processes remained.
| Panel | Complete pairs | Prior median external wall | Current median external wall | Median paired current/prior ratio | Geometric mean of per-case paired median ratios |
|---|---|---|---|---|---|
| 33 common FAST cases | 5 | 5.6303 s | 5.7401 s | 1.0195 | 1.0024 |
| Five existing MIPLIB inputs | 3 | 1.7816 s | 1.5935 s | 0.9421 | 0.9608 |
The ten complete FAST panels passed 330/330 case executions. The six MIPLIB panels passed 30/30 solves with independent exact witness and known-optimum checks on p0033, p0201, p0282, p0548 and lseu. No case triggered the predeclared diagnostic threshold: its median must slow by both more than 20% and more than 1 ms, and the joint condition must recur in at least half its paired runs. This is a diagnostic rule, not a statistical significance test.
One additional attempted panel failed its deadline. Six FAST pairs were planned; the current-library panel starting pair six timed out in cp_distinct at 4.0046 s under the existing four-second cap. Its other 32 cases passed. The failure is preserved, was not retried, and is excluded from the five complete paired medians. The final comparison explicitly records all_attempted_fast_panels_passed: false. Successful runs of that case ranged from 3.29 to 3.55 s. Its executable and all nine native-library instruction sections are byte-identical between cohorts. This shows a deadline-margin and repeatability limitation in unchanged code; it does not identify the cause of the timing variability. No limit was relaxed to obtain a pass.
The common FAST results suggest little change on these workloads. The MIP panel's lower measured times do not establish a general speedup. This short comparison includes process startup, loading, polling and desktop noise; it has no large-instance holdout, confidence intervals or memory analysis. It compares with the preceding integrated checkpoint, not CPLEX/Gurobi or the entire original benchmark corpus. New-feature correctness is covered by the separate 35-case gate; this comparison does not measure new-feature performance benefits.
Full raw measurements, per-case times, provenance and the failed sixth panel are in build/comparison-eighth/results-final/. The interpretation is interpretation.md; the full machine-readable record is comparison.json. The preserved comparison script and preparation manifest are in its parent directory. A committed compact record is checkpoints/ninth.json.
Usable artifacts and boundaries
The combined executable is build/native-compat/bin/fzn-gecode-optimize and the combined facade is under build/native-compat/gecode/optimize/. Fresh tested installation prefixes are build/stage/combined-ninth, build/stage/standalone-ninth and build/stage/core-ninth. The build/API guide, MiniZinc guide and FAST command describe their use. Exact executed FAST arguments and library hashes are in build/optimize/ninth-fast-wall.json and ninth-fast.json. Existing eighth-checkpoint stages and binaries were preserved.
This batch did not rebuild the Python wheel; the existing eighth-checkpoint wheel does not contain these additions. No package publication, remote CI, actual Windows/Linux runtime validation or oldest-macOS execution is claimed. The full native sanitizer configuration instruments the native libraries, HiGHS and facade; leak detection is disabled, with ASan/UBSan halt-on-error enabled. No further branching, learning, nonlinear or parallel-search work is started by this checkpoint. The broader parity roadmap remains incomplete.
Generated by
1.18.0