Setup Test is failing

[1-Bort-1]

TL;DR

The audit plot_slices_3d crashed on a generated deflection .dat that reads back with no coordinates, taking the whole Setup Test down with ram_air_kite.jl; it now skips that overlay and warns, naming the file and whether it was blank or absent.

What was wrong

ram_air_kite.jl:79 runs the audit plot over the generated polars_xfoil directory at delta=1.0. generated_slices read airfoils/<id>_d1.dat, got zero points back, and handed the empty contour to the Kulfan fit: ArgumentError: reducing over an empty collection out of argmin in normalize_airfoil (src/airfoil_aero/kulfan.jl:71), through fit_pts at ext/VortexStepMethodMakieExt.jl:1676. The file was there, so the isfile branch that already warns and draws nothing for a deflection that was never generated did not fire.

It reads back empty because it holds NaN NaN rows and read_dat_coordinates drops any line that does not start with a digit, - or .. A deflection column of SectionAero.x stays at its fill(NaN, …) initial value when no angle of that deflection lands a solution of the grid's modal node count (src/airfoil_aero/section_aero_gen.jl:27-38), and write_section_aero writes that column out unguarded — three lines above the guard that already refuses to write {id}.dat all-NaN for exactly this reason.

So the audit plot has two ways to have no deflected contour to draw and only handled one of them. It now handles both the same way, and the warning says which: Skipping the delta=1.0° overlay for 4_d1.dat: no finite coordinates; … against … 4_d1.dat: no such file; ….

Is the blank column a regression in the wrapping under delta deflection?

Asked on the thread, and worth the check — a plot that skips a contour is a plot that can hide a bad one. I measured the deflected wrap directly rather than through the plot, on the example's own geometry and settings (10 slices of ram_air_kite_body.obj, wingtip_distance=0.1, ShrinkWrap(clearance=0.0), crease_frac=0.75, Re=1e6, XFoil npan=160/ncrit=9/xtrip=(0.05, 0.05), α = −8:2:26°).

Sections 1 and 10 are thrown out on thickness (0.424 against a 0.225 limit) before any deflection is formed — that is the Reused the nearest valid airfoil for: section 1 → airfoil 2, section 10 → airfoil 9 line in the CI log, and it is about the raw tip slices, not about δ. For the eight sections that remain, the 1° deflected wrap is indistinguishable in health from the undeflected one: 239 contour nodes either way, no non-finite coordinate, chord normalised to [0, 1], no fold on either surface, enclosed area within 0.1% of δ=0, and XFoil converges at exactly the same 16–17 of the 18 angles. No blank column reproduces on this box at all.

That matches what the mechanism says it should be. The contour that reaches the .dat is deform_section's own output — #301 measured XFoil handing it back unchanged at repanel=false, max|dx| = max|dy| = 0.0 — so the shape never depends on the solver converging; only its presence does. A blank column therefore means "no angle of that deflection landed", not "the deflected wrap produced garbage". generate_airfoils validates only sols[1], the δ=0 column (src/airfoil_aero/geometry_gen.jl:48), so a dead deflection column leaves no trace anywhere else.

Per-section numbers, δ=0 against δ=1°

sliced 10 sections; thickness limit 0.2253; degenerate: [1, 10]
section 2:  d=0  n=239 nonfinite=0 chord=[0.0000,1.0000] tmax=0.1155 area=0.12906 folds=0/0  converged 16/18
            d=1  n=239 nonfinite=0 chord=[0.0000,1.0000] tmax=0.1165 area=0.12915 folds=0/0  converged 16/18
section 3:  d=0  n=239 area=0.13949 folds=0/0  converged 16/18      d=1  n=239 area=0.13962 folds=0/0  converged 17/18
section 4:  d=0  n=239 area=0.13247 folds=0/0  converged 16/18      d=1  n=239 area=0.13256 folds=0/0  converged 16/18
section 5:  d=0  n=239 area=0.13016 folds=0/0  converged 17/18      d=1  n=239 area=0.13026 folds=0/0  converged 17/18
section 6:  d=0  n=239 area=0.13018 folds=0/0  converged 16/18      d=1  n=239 area=0.13028 folds=0/0  converged 16/18
section 7:  d=0  n=239 area=0.13249 folds=0/0  converged 16/18      d=1  n=239 area=0.13259 folds=0/0  converged 16/18
section 8:  d=0  n=239 area=0.13955 folds=0/0  converged 16/18      d=1  n=239 area=0.13967 folds=0/0  converged 16/18
section 9:  d=0  n=239 area=0.12914 folds=0/0  converged 17/18      d=1  n=239 area=0.12924 folds=0/0  converged 17/18

What this change can and cannot hide

It cannot hide a wrong deflected contour. The skip fires only on a contour with no usable coordinate at all; anything finite is fitted and drawn exactly as before, and the fit_pts guard returns an empty point vector only for an input that previously threw. The deflected contour is also the only one that can reach it empty: write_section_aero already refuses to write {id}.dat all-NaN, and the live path's shrink_wrap throws rather than hand back nothing — so no overlay goes quiet without the warning above. What it does change is that a deflection the solver never landed now reaches the reader as a named warning rather than as a stack trace — and, after this round, as a warning that does not read like a missing file, which is the case #298 is about.

Found on the way

deform_section's docstring (src/airfoil_aero/airfoil_solvers/common.jl:100-102) still promises that the re-wrap "bridges the crease with a min_concave_radius fillet". Since 1e13d92 the wrap offsets the ball's contact polygon instead of sweeping its rim, so a concavity narrower than the ball is bridged straight (src/airfoil_aero/shrink_wrap.jl:139-146), and at the clearance = 0.0 that path uses, the contour is the contact polygon itself. The crease geometry the solvers are handed is therefore not what that docstring says. It is behaviour-neutral to fix and in a file this diff has no other reason to open, so it is not in here.

The producer-side half — that write_section_aero writes the all-NaN column at all, and that generate_airfoils never notices a deflection that converged nowhere — is #300, open as #301. The two are complementary: with #301 merged the file is absent rather than blank and this plot takes its other branch, and until then (or for a directory generated before it) this is what stops the crash.

Verification

• Reproduced first: ArgumentError: reducing over an empty collection from normalize_airfoil/argmin, ext/VortexStepMethodMakieExt.jl:1632 → kulfan.jl:71, identical to the Setup Test stack on run 34653119608
• test/plotting/test_plotting.jl "Audit slices (Makie)" red before, green after (juliaserver): the empty-contour assertions errored before 4c799ad; the warning assertion failed before 8ded002 with the old text No generated .dat for delta=1.0° (4_d1.dat); now 7/7, file 84/84
• Local full suite: PASS (7 min) on 8ded002 · GitHub CI on 8ded002: Setup Test green, CI green on the rerun
• The one red job on the first CI attempt was test/solver/test_forwarddiff.jl's AD-against-finite-difference check, relative_error = 0.0400 against 1e-4, with norm_fwd = 2.114327230856032 — The generated ram-air polar fixture is not reproducible, so every test asserting numbers on it draws a fresh geometry each run #291's documented draw of the regenerated ram-air fixture, in code this diff does not touch (no src/ change at all); green on rerun of that job alone
• Docs build: not run — no docstring, export or docs/src page changed
• Benchmark: n/a
• Risk: the blank column is a solver outcome that does not reproduce on this box, so the crash path itself is exercised only by the regression test's synthetic blank .dat, not by a generated one.

Scope

+48 / −10 across 3 files: ext/VortexStepMethodMakieExt.jl (the skip and the warning), test/plotting/test_plotting.jl (the regression test), CHANGELOG.md. No stack; #301 touches the writer and the readers, not this.

Closes #311 · task VortexStepMethod.jl-311

[1-Bort-1]

Local full suite: PASS (6 min, Julia 1.12.7, one cell of the matrix)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[1-Bart-1]

Are you sure this is not a regression for the wrapping under delta deflection?

[1-Bort-1]

No — I measured the deflected wrap itself rather than through the plot. At δ=1° all eight non-degenerate sections of the example's own geometry give 239 contour nodes (as at δ=0), no non-finite point, no fold, enclosed area within 0.1% of δ=0, and XFoil converges at the same 16–17 of 18 angles; sections 1 and 10 are dropped on thickness before any deflection is formed. The blank _d1.dat is a deflection that landed no angle, not a bad wrap: the contour written there is deform_section's own output, which XFoil hands back unchanged. Per-section numbers are in the PR description. The skip also could not hide a wrong wrap — it fires only on a contour with no usable coordinate — and 8ded002 makes the warning say which file was blank instead of reading as a missing one.