Fix NONLIN solver returning stale gamma on repeat solve! calls

[1-Bart-1]

Summary

The NONLIN solver path returned stale gamma on every solve! call after the first. The cached NewtonRaphson SciML cache was reused without reinit!, and NonlinearSolve short-circuits caches that already terminated with ReturnCode.Success, so the second call onward did no work — even when geometry, AIC, and inflow had all changed. Effectively only the first solve on a given solver was real.

What changed

• Replaced the SciML cache with a hand-rolled Newton. SciML reinit! cascades through inner LinearSolve/Jacobian caches and allocates ~250 k per call, breaking the bench allocation budget. The hand-rolled loop uses pre-allocated buffers and LinearAlgebra.LAPACK.getrf! / getrs! for the LU step, all in-place. Re-solves correctly on every call, no allocations in the hot path.
• Convergence check rewritten to SciML-style on the Newton increment: ‖Δγ‖_∞ < atol + rtol · max(‖γ‖_∞, tol_reference_error). rtol now has the same meaning as in the LOOP solver.
• linearize returns a third value converged::Bool that is true iff every internal solve reached the tolerance; emits a warning when not. Existing callers using jac, res = linearize(...) keep working (Julia destructures the head of the tuple).
• linearize gained fd_absstep / fd_relstep kwargs for the inner FiniteDiff Jacobian step.

Tests

• test/solver/test_solver.jl — regression test that runs two NONLIN solves with different va and asserts gamma actually changes (was bit-identical pre-fix).
• test/body_aerodynamics/test_results.jl — replaced the old per-input error_ratio < 0.002 checks with a two-scale Jacobian validation: at scale=1 the predicted change must match the observed change to FD precision (sanity), at scale=2 it must match within an O(h²)-dominated tolerance (verifies the Jacobian reflects true sensitivity, not noise). Also asserts lin_converged.
• test/polars/test_polars.jl — populated with an XFoil polar generation test (small range, ~3 s) so that path is still covered now that the linearize test no longer regenerates polars.
• The cached ram_air_kite polar CSVs are committed as test fixtures (42 KB total) and removed from .gitignore. Without them, CI was generating tiny 3×3 polars from alpha_range=deg2rad.(-1:1) that did not cover the actual flow regime, causing Newton to fail.

Test plan

• julia --project=. -e 'using Pkg; Pkg.test()' — full suite green
• Bench allocation budget respected (no 250 k regression from SciML reinit!)

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 98.07692% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/solver.jl	98.07%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[Copilot]

Pull request overview

Fixes a correctness issue where the NONLIN solver could return a stale circulation (gamma) on subsequent solve! calls, and adds regression coverage for repeated solves on the same Solver instance.

Changes:

• Adds a regression test ensuring two consecutive NONLIN solve! calls with different inflow (va) produce different gamma.
• Reworks the NONLIN path in gamma_loop! to use an in-house Newton iteration with finite-difference Jacobian + in-place LU solve.
• Adjusts nonlinear-vs-linear test tolerances and marks several linearization accuracy checks as @test_broken.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 6 comments.

File	Description
src/solver.jl	Replaces the previous SciML-based NONLIN implementation with a custom Newton + LU-based solver loop.
test/solver/test_solver.jl	Adds regression coverage for repeat NONLIN solves on the same solver instance.
test/body_aerodynamics/test_results.jl	Loosens tolerances and marks some linearization validation assertions as broken.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.