[fhenskens]

Solver buffer allocation, compact-model work, and repeated derivative evaluations add overhead to bounded quasi-Newton solves. This PR reduces that work and adds opt-in L-BFGS-B model strategies and projected L-BFGS searches. Existing defaults and checkpoint layouts are preserved; no dependency, version, or MSRV changes are included.

Changes

• Reuse TNC/L-BFGS-B buffers and compact-model storage, retain accepted Wolfe evaluations, and add opt-in fused Wolfe evaluations for combined cost/gradient callbacks.
• Reduce sequential finite-difference Jacobian allocations and single-worker Jacobian/Hessian scheduling overhead.
• Add signed-column/scaled-compact L-BFGS-B models with bounded, interior-first, and projected-chord direction policies. ModelDirection::ProjectedChord is the single chord-policy name.
• Add explicit feasible-ray, projected Armijo, and endpoint-chord searches through with_projected_search(). Form curvature pairs from actual accepted displacements, support retention or reset-on-mask-change, and serialize the preceding mask for exact continuation.
• Document configuration, backend support, source attribution, callback accounting, and numerical-policy differences; add analytic, backend, checkpoint, and reference coverage.

Recorded performance

These figures were checked by recomputing summaries from the latest recorded raw runs on November 5. Baseline: Basin a673d9f. These are native f64/Vec measurements with combined callbacks.

Ratios are median previous default time / method time over common target successes, shown as cheap / synthetic-expensive callbacks. Hits mean the first observed normalized objective gap <= 1e-8, including rejected probes, and do not imply native convergence or returned-point accuracy. Misses and failures remain in the success denominators.

Projected search, retention enabled	Basin targets / 206	Speed vs baseline (cheap/expensive callbacks)
Feasible ray	171	1.560x / 1.078x
Projected Armijo	167	1.582x / 1.041x
Endpoint chord	170	1.654x / 1.130x

Default current and baseline L-BFGS-B both reach 170/206 targets with identical recorded numerical results and callback work. Projected methods disable relative-cost stopping while the baseline retains it: these ratios measure solver effectiveness under different policies, rather than isolated implementation speedups.

Retention shows no universal benefit: reset-on-mask-change reaches 172 ray and 171 chord targets, versus retained 171 and 170; both Armijo settings reach 167. Median callback ratios are one and timing differences are within about one percent. All projected settings and references reach 7/8 targets on the additional starts. Small timing differences are inconclusive.

The latest final TNC run records 194/206 observed targets for baseline and ordinary/fused current Basin. Fused TNC returns 191/206 points meeting the target. Its target-time ratios are 1.200x / 1.373x versus baseline, with a median target-callback ratio of 0.714x versus baseline. Fusion remains opt-in; these figures apply to combined-callback workloads.

Benchmark protocol and audit

The projected-search run contains 28,248 measured solves: 206 scenarios plus eight additional starts, two callback costs, separate total-only/target-logging passes, and three repeats, excluding warmups. The earlier final TNC/model run contains 30,816 solves. Reanalysis of both raw JSONL sets reproduces their stored summaries exactly. Recorded manifest and projected-run executable hashes were checked.

All contestants use Rust 1.99.1, thin LTO and one codegen unit, fixed dependency locks, history capacity 10, a 20,000 raw-callback guard, and a 5,000-iteration limit. Thread counts are one; CPU affinity is not pinned. Expensive callbacks add 2,000 synthetic work units. Baseline/current process order alternates; candidate and reference methods are shuffled together.

Projected methods use projected-gradient tolerance 1e-6, gradient-only stopping, and relative global curvature tolerance 1e-10. Armijo uses 40 trials, c1=1e-4 and contraction 0.5. More-Thuente uses 20 trials, c1=1e-3, c2=0.9 and relative bracket tolerance 0.1. Basin retains its rounding/recovery safeguards. These previously inspected scenarios provide regression evidence, not unseen generalization tests or confidence intervals.

The audit checks repeat/pass consistency, finite feasible returned points, and callback guards. A freshly rebuilt final-source harness also reproduces all numerical and work fields across 4,280 cheap-callback records (206 scenarios plus eight starts, all ten methods and both passes); its timings are not substituted into the recorded performance comparisons. Published website benchmark data are unchanged.

Validation

• Routine Rust suite: 2,858 passed, one existing ignored test. The known baseline NEWUOA chained_rosenbrock_6d failure was excluded and separately reproduced on the baseline.
• Rechecked all three formatting scopes and all 27 compact-model/projected-search integration tests across every dense backend version, f32/f64, and serde, including geometry, retention, exact postcard continuation, and fresh dimension reuse.
• Analytic boundary/non-finite/error/accounting checks and dense BFGS/Cauchy and Fortran/reference-driver checks pass.
• Core all-target/all-feature Clippy, warnings-denied public rustdoc, guide examples, direct Markdown/Svelte compilation, and both default/no-default WASM builds pass.
• Full workspace Clippy is blocked locally by missing libclang for the external C-reference dependency. Full web checks are blocked by missing Biome/wasm-pack/generated WASM and Windows mdsvex paths; direct compilation of the changed guide pages passed. CI must supply full workspace and site checks.

[jolars]

Thanks! Looks good, just one small nit.

[jolars]

When callers use Executor::from_start, selecting .with_compact_model() fails with E0277 because the returned ModelLbfgsb does not implement InitialState. The existing Lbfgs supports this constructor, and the wrapper still needs only a starting point. Forward InitialState<V, State = FirstOrderState<V, F>>, as already done for ProjectedLbfgs, so callers can select the compact model without constructing the state manually.

[fhenskens]

Oh right. My mistake, I'll fix that up. Cheers!

[jolars]

Btw, can you check the website failure? I see:

 * [404] <https://github.com/fhenskens/sisyfuzz/tree/39f6d4d0062607eb48b07d5cd5f847fb7da13828/src> (at 283:15) | Rejected status code: 404 Not Found

Maybe you can either replace that URL with a permanent link or add it as an exception to the url checker.

[fhenskens]

For sure. Forgot to make it public. Don't need the link in there anyhow. Your project is a much better home for the work. Thank you.

[jolars]

Thanks for the fixes!