[1-Bort-1]

TL;DR

New test/verification/test_verification.jl ports the five cases in the Python tests/verification_cases that check against something outside the code: an elliptic wing against lifting-line theory, a curved Clark Y wing against RANS, a rectangular AR 12 wing against CFD, the V3 kite against RANS, and three horseshoe vortices against Biot-Savart. Until now nothing in the Julia suite compared a whole-wing CL or CD with theory or measured data. The existing reference tests only compare against the old thesis port, which is a copy of the same maths.

What was ported and what was not

Each case keeps the Python geometry, flow speed, relaxation factor, core_radius_fraction = 1e-20 and angles of attack, and every tolerance is copied unchanged from awegroup/Vortex-Step-Method@c82d521. The Python asserts comparing its "old" thesis code with its "new" package are left out, because they compare two Python codes. Geometry reuses the generators already in test/utils.jl. The V3 kite's 10 CAD section pairs are written into the test and split 4× with SPLIT_PROVIDED (37 sections, as in Python). The six CSVs are in test/verification/data/, with a README giving the source of each.

Julia passes every tolerance. The tightest margins are the V3 kite's CL at 9° (0.079 against 0.1) and the elliptic wing's VSM CD (0.00137 against 0.002):

case	reference	Julia VSM	Julia LLT	tolerance
elliptic AR 20, 3°/9°	CL 0.2991/0.8972, CDi 0.001424/0.01281	CL 0.2954/0.8861, CD 0.001578/0.01418	CL 0.2991/0.8967, CD 0.001423/0.01278	LLT CL 1e-2, CD 1e-3 · VSM CL 1e-1, CD 2e-3
curved Clark Y, 9°	CL 0.7921, CD 0.07507	CL 0.7904, CD 0.1034	CL 0.9093, CD 0.1029	VSM 1e-1 · LLT−VSM CL 2e-1, CD 4e-2
rectangular AR 12, 3°/6°/9°	CL 0.6094/0.8821/1.144	CL 0.6252/0.8618/1.101	CL 0.6374/0.8868/1.141	VSM CL 1e-1 · LLT−VSM CL 2e-1, CD 4e-2
V3 kite, 3°/6°/9°	CL 0.4608/0.6728/0.8605, CD 0.04815/0.06400/0.08148	CL 0.4470/0.7020/0.9391, CD 0.0685/0.0960/0.1346	–	VSM 1e-1

Where the port differs from the Python test

• Horseshoes: the Python test sets the bound-vortex velocity to zero whenever the evaluation point has the filament's x (if evaluation_point[0] == filament.x1[0]), so it only ever tested the trailing legs. The Julia test adds all three filaments. The bound filament really does induce exactly zero here, because the evaluation point lies on its line, so the reference values are unchanged.
• "Swept" wing: Python's swept_wing case has sweep 0. It is named for what it is here: a rectangular AR 12 wing.
• Clark Y polar: Python's clarky_maneia.csv repeats its 20–32° block three times, which makes alpha non-monotonic, and its comment says radians although the values are degrees. Only the first block is kept.
• Kite CD reference: Python first resamples the CD data onto the CL data's angles and then interpolates again. Here CD is interpolated directly, which moves the reference by at most 3e-4 against a 0.1 tolerance.

Found on the way

• .github/workflows/CI.yml:22 has fail-fast: false, which the agent rules want gone. I did not change it here, since it is a separate cleanup: change.
• The drag tolerances are loose because upstream widened them in awegroup/Vortex-Step-Method@fac126b, when its VSM stopped moving alpha from 3/4 chord to 1/4 chord. Julia's VSM CD is 0.03–0.05 above the RANS values on the curved wing and on the kite at 6–9°. That is the drag alpha is built from the full relative velocity, not its component perpendicular to the quarter-chord line #317 and correct_aoa defaults to false, so the solver runs as LL-3/4 and loses 24-59% of the induced drag #318 are about, and these tests will still pass after either fix.

Verification

• Reproduced first: n/a, these are tests of existing behaviour with no bug to reproduce
• The tests can fail: making the semi-infinite trailing vortex 10 % stronger in src/filament.jl gave 18 passed, 4 failed (elliptic VSM CD and the three semi-infinite horseshoe sums). With src/ restored it went back to 22/22 green
• test/verification/test_verification.jl: 22/22 pass. Through runtests.jl together with body_aerodynamics/test_body_aerodynamics.jl, which includes the same utils.jl: 4880/4880 pass, verification set 18.7 s cold (juliaserver)
• Local full suite (agent ci-local, Julia 1.12): still running when this was written · GitHub CI: pending
• Docs: n/a, no public symbols · Changelog: none, test-only · up to date with main
• Risk: the kite CL at 9° has 0.021 of headroom, so a physics fix that raises kite lift may need this reference looked at again, not the tolerance loosened.

Scope

+350 / −0 across 10 files, all under test/. That is 226 lines of test (50 of them the kite's CAD coordinates), 107 lines of reference CSV, a 13-line data README, and one line each in runtests.jl and test/README.md.

Closes #93 · task VortexStepMethod.jl-93

[1-Bort-1]

Independent review (advisory)

Verdict: APPROVE WITH COMMENTS · 2 inline, 0 off the diff

Good

• This is what #93 asked for: the five Python cases that check against theory, CFD, RANS or Biot-Savart, and no Python-vs-Python checks; I compared the diff with the card and every testset maps to a card row
• Geometry comes from the existing generate_coordinates_* and flip_created_coord_in_pairs in test/utils.jl; the new helper names (read_columns, polar_vectors, max_error, …) collide with nothing else in test/, which I checked with grep
• The card's margins match its table: kite CL at 9° is 0.9391−0.8605=0.079 and elliptic VSM CD is 0.01418−0.01281=0.00137
• The card shows the tests can fail: a 10 % stronger semi-infinite vortex turned 4 of the 22 checks red
• Reusing left_velocity/right_velocity in the finite-trailing check is safe: velocity_3D_trailing_vortex! overwrites vel in both branches (src/filament.jl)
• No line in test_verification.jl is over 92 characters (checked by grep); the DelimitedFiles and Interpolations dependencies are already in test/Project.toml
• The card explains each departure from Python (the sweep-0 name, the tripled Clark Y block, the direct CD interpolation) with its size against the tolerance

Not good

• test/verification/test_verification.jl:39 — wing_from_coordinates is the same flip → Wing(n; UNCHANGED) → add_section! pairs → refine! loop written out twice in test/body_aerodynamics/test_body_aerodynamics.jl (lines ~32 and ~172), which includes the same utils.jl. Adding it only here makes a third copy; it belongs in test/utils.jl, with the body-aero copies switched to it now or in a cleanup: PR.
• test/verification/test_verification.jl:201 — In all three horseshoes the evaluation point lies on the bound filament's line, so bound_velocity is exactly zero and the change sold as 'adds all three filaments' checks nothing Python did not. Either assert bound_velocity == zeros(3) so the claim is tested, or drop the call and say in the card that the case only covers the trailing legs.
• The card does not say what was searched for before the new helpers were added (§2), and the duplication below suggests the search missed test_body_aerodynamics.jl
• The local full suite and GitHub CI were still unticked when the card was written, so the only green result is from the verification file plus body_aerodynamics
• Kite CL at 9° has 0.021 of headroom; the card rightly says to revisit the reference rather than widen the tolerance, and that should hold when #317/#318 land
• max_error is the only helper without a docstring; its name is clear, but the file is otherwise consistent about docstrings

_{claude, rubric CLEAN_CODE.md. A different lab from the implementer
on purpose: a reviewer sharing its blind spots would not flag its mistakes.}

[1-Bort-1]

MINOR: wing_from_coordinates is the same flip → Wing(n; UNCHANGED) → add_section! pairs → refine! loop written out twice in test/body_aerodynamics/test_body_aerodynamics.jl (lines ~32 and ~172), which includes the same utils.jl. Adding it only here makes a third copy; it belongs in test/utils.jl, with the body-aero copies switched to it now or in a cleanup: PR.

[1-Bort-1]

MINOR: In all three horseshoes the evaluation point lies on the bound filament's line, so bound_velocity is exactly zero and the change sold as 'adds all three filaments' checks nothing Python did not. Either assert bound_velocity == zeros(3) so the claim is tested, or drop the call and say in the card that the case only covers the trailing legs.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!