[1-Bort-1]

Julia was translated from Python v1.0.0, and Python's v2.3.0 changed its default results. On swept and curved wings, which is every kite, the two had drifted several percent apart. With this branch, Julia matches Python's circulation, angles, per-panel forces and moments, totals and coefficients to six digits on nine reference cases generated by running the Python code, with both solved to a tight stop. At the default stop, circulation, angles and coefficients stay within 1e-5 of those results.

wing, 8°	CL before	CL now = Python	CD before	CD now = Python
straight	0.6906	0.6910	0.0186	0.0151
swept, tapered	0.8456	0.8256	0.0364	0.0299
arc	0.5947	0.5987	0.0354	0.0325
V3 kite	1.0509	1.0992	0.4910	0.4575

What moved: circulation now follows the effective speed. The chord and chordwise axis are measured square to the bound vortex. Forces turn with the flow at the quarter chord, including a correction for a curved bound vortex, and drag runs along that flow. The force on the chordwise legs is on by default and shared per section boundary. Each wake follows its own panel's inflow, and reference pressure and directions leave out body rates. The defaults also change to 5000 iterations and a relaxation factor of 0.8 of the stability limit, read off the polar the way Python does, so the curved reference wing that used to diverge now converges.

The same default change ends the lost induced drag of #318: the straight elliptic wing of aspect ratio 4 from Gaunaa, Li & Pirrung 2026 now reaches an Oswald efficiency of 0.9947, against 0.7295 with the quarter-chord force direction off.

Agreeing with Python does not make swept wings right. Against that paper's vortex-lattice results, Oswald efficiency is off by −8.2 % at 15 % sweep, −8.0 % at 10 % sweep with 25 % anhedral and −11.1 % at 25 % sweep, where the solver does not converge. The paper's own model is off by +5.6 %, −0.5 % and +1.1 %. Purely curved wings are off by ±2.4 %. These were measured on the previous head by the task behind #400, whose card plots them, and the chord direction on the swept wings is an assumption. #416 tracks the gap; closing it may mean leaving Python's rule, which is the owner's choice.

This is a breaking change even where the results are more correct: the same input now gives different numbers, and the changelog marks it so. That reaches KiteModels wherever its settings do not name these options. Setting the two force options to false and naming a relaxation factor restores those switches, but not the geometry or circulation changes.

Decisions I made:

• A semi-infinite filament's zero on its own axis is not marked breaking, because the NaN before was no usable result. The standing rule on breaking changes is proposed in 1-Bart-1/Agents#902.
• Convergence is still measured on the unrelaxed residual at 1e-5, because it does not depend on the relaxation factor. Python's relaxed step at 1e-6 is similar in size on most wings and looser on the V3 kite. Julia's stop leaves circulation within 5.5e-6 and cl and cd within 5.4e-6 of the converged result.
• Warm starting from the previous circulation stays on, because simulation loops rely on it. Past stall this can pick a different solution from Python, which starts from zero.
• Sideslip from angles keeps Julia's unit-length apparent wind. Python's is not unit length (1.00045 at 10° and 10°), so the reference cases pass the inflow as a vector.
• With several wings the curved correction acts only within each wing; Python refuses more than one wing.
• The correction is kept per shape and rebuilt after any shape change. Python's reuse under rigid rotation is not ported.
• The correction is not rebuilt when the inflow speed changes, because that would cost a rebuild per simulation step. Speed only sets the size of the trailing core, and rebuilds at 1, 10 and 100 m/s are identical to the last bit on all nine cases. A point that did fall inside a core would keep the speed of the last rebuild.
• Drag beyond a polar table is now held, as in Python, for vector polars only. Matrix polars have no Python counterpart and are unchanged.
• Inside a trailing vortex's core the speed is no longer limited, a deliberate match to Python: closer than 7 micrometres to the line at 10 m/s, the speed is the value on the core boundary, 1.85e5 m/s at 1e-7 m for unit circulation. The nine reference cases still match.

Found on the way: faster convergence exposed that unrefined section coefficients were plain means, so they are now area-weighted, with the same code as #402. In that code the section chord becomes area over width, and the section moment becomes the sum of its panels rather than their mean. That moment, and the linearize outputs and Jacobian rows taken from it, now grow with the panels per section. A semi-infinite filament now gives zero rather than NaN on its own axis, as #408 also does. It also counts a point within rounding of its line as on the line, by the rule the finite filament uses; such a point got 1.85e9 m/s at 1e-15 m.

This covers default results only, so the issue stays open. The missing features are queued as parts of it, each with its Python tests. Of Python's 198 tests, 84 are pinned in Julia and #417 ports 14 more whose feature already exists; the rest wait on those parts, were declined, or test Python-only mechanics such as its compiler and CasADi. That count comes from reading both suites, not from a run. Left out: bridle drag, which stays in KiteModels as ruled on #113, and the other items declined there (flat area, rotate, cosine circulation, configurable aerodynamic-centre fractions) unless the owner says otherwise.

To act on: stacked on #392, which merges first. It takes a different route from #400, whose rule is not Python's, and from #408's core scaling. It conflicts with #402 in the solver, where the aggregation code is identical. SymbolicAWEModels calls the drag-direction helper with an argument that is now gone, so it needs a one-line change when it raises its bound to the release #392 forces.

Verification

• Reproduced first, unchanged code against Python: V3 CL 1.0509 vs 1.0992, CD 0.4910 vs 0.4575, chord off by up to 0.33 m
• test/verification/test_python_reference.jl: green on nine cases (straight, swept, arc, sideslip, LLT, viscous, body rates, distributed inflow, V3). Red before: run through a script on the unchanged code, since the test itself names functions this branch adds
• test/verification/test_verification.jl: the aspect-ratio-4 test is green, 24/24 in the file; red with the quarter-chord force direction disabled (0.7295, 23/24)
• Green: 20 affected test files, including test/bench.jl, test_verification.jl, test_curved_bound_correction.jl, test_forwarddiff.jl; the six a later filament change touches (test_bound_filament.jl, test_panel.jl, test_curved_bound_correction.jl, test_python_reference.jl, test_forwarddiff.jl, test/bench.jl) rerun on the previous head; on this head the ten files the last fixes touch, and the docs
• Docs build clean in a fresh session · repo has no REUSE lint
• Local full suite: running on this head (Julia 1.13.3, one cell of the matrix) · GitHub CI: this draft runs the Documentation job only; the Julia matrix starts once it leaves draft
• Main's four newer commits, merged temporarily and backed out: test_wing_geometry.jl, test_unrefined_dist.jl and test_python_reference.jl green, no conflict
• test_python_reference.jl: the default stop and default iteration limit converge within 1e-5 on all nine cases, and the relaxation factor matches Python to 1e-12, red before on V3 (0.0214 vs 0.0232). A repeat solve stays under 1 kB, red at 4240 B on a first version of that fix. test_semi_infinite_filament.jl: rounding off the line gives zero, red before at three distances. test_forwarddiff.jl: the stability limit is type-stable on dual numbers, red before (Union{Float64, Dual})
• test_curved_bound_correction.jl: rebuilding the correction allocates under 2 kB, red before the fix, when a V3 solve after a shape change allocated 1,536,208 B at 18 panels
• Allocations, V3 at 18 panels: a solve after a shape change 1744 B, against 208 B with the two force options off; a repeat solve 208 B either way; calc_forces! 0 B
• Iterations from a cold start, new defaults against the two force options off at the old relaxation factor 0.03: straight 46 vs 322, swept 55 vs 294, arc 39 vs 300, V3 291 vs 223
• Benchmark: not run. The box stayed above its load limit for the full 8-hour wait (load 34 of 20 cores at the end), so no timing is claimed here
• Added with no red run, since they pin behaviour that already held: the closed-form limit, a named factor, a new inflow keeping the correction, a panel in still air keeping its wake, and the moment of the leg forces.
• Tests that follow the change, same system: the two rectangular-wing coefficient literals had moved inside their tolerance and now carry the new values, CFz tightened to 1e-4. thesis geometry expectations take the square chord and frame, same tolerance (Python loosened its own to 1e-3). Pinned rectangular-wing values move 0.1 % (force x −117.965 → −117.838) with the circulation map. Tests that compare against section-only formulas now set the leg force off. cl is checked against the control-point angle. The NaN assertion on the filament axis is now zero. One trim call tightens its angle tolerance to 1e-6, since its 1e-5 assertion sat at 1.17e-5.
• Reference: Python a2a91cb, numba 0.68.1, numpy 2.5.4; regenerate with test/data/python_reference/generate_reference.py
• Risk: simulation loops that change the shape every step now rebuild the correction every step, 148 filament evaluations per panel pair (47,952 at 18 panels). A solve without a named relaxation factor also makes lift lookups to find the limit, without allocating: two per table angle per panel (756 for V3 at 18 panels), or 182 per panel for polars without a vector table. Neither time is measured.

Closes #318

Opened by 1-Bort-1, an AI agent working for @1-Bart-1.
Refs #410 · task VortexStepMethod.jl-410

[1-Bart-1]

More correct results but changed. So this is breaking change.

[1-Bart-1]

Also this reads as "we edited the results" I want to see, "we changed this to agree with the python version and the gaunaa paper and now it agrees with the python version"

[1-Bort-1]

Marked BREAKING in 2e85236: default results, solver defaults, unrefined distributions, section moment and LLT correct_aoa; also added to the standing rules.

[1-Bort-1]

Rewritten in dcffb9e: the entry now says the solver changed to agree with Python v2.3.0 and Gaunaa 2026, and that results now agree with Python, with swept wings still off the paper (#416).

[1-Bart-1]

And did we port all python tests as well? And update all tests?

[1-Bort-1]

No to the first: not all ported yet. Of Python's 198 tests, by reading both suites, 84 are pinned in Julia and 16 had a feature but no test. 0cab56c takes the 2 of those on this branch's own code; #417, which I should have queued earlier, ports the other 14. 48 wait on missing features (#411 16, #412 8, #414 3, Masure on #387 10, declined on #113 10, zero-speed wake 1) and 50 test Python-only mechanics like numba and CasADi.

Mostly yes to the second, with holes now fixed in 0cab56c. No assertion was loosened and the suite passed, but the reference test skipped V3's relaxation check, which hid a real mismatch (0.0214 vs Python's 0.0232, now equal). Two coefficient literals had drifted inside their tolerance, nothing ran the default iteration limit, and the leg force's moment check was gone. The Julia matrix has not run, because this PR is a draft; the local suite on 0cab56c is running.