(compat): Upgrade FastInterpolations to v0.4 and AdaptiveArrayPools to v0.3

[mgyoo86]

Summary

Migrate all interpolation call sites across the codebase to the FastInterpolations v0.4 API and bump AdaptiveArrayPools to v0.3. This is a mechanical but codebase-wide change (22 files, all modules).

FastInterpolations v0.4

Full migration guides:
📖 v0.2->v0.3
📖 v0.3->v0.4

Key changes reflected in this PR:

• Series() wrapper required for multi-column (series) interpolants:

  # Before
  cubic_interp(xs, matrix; bc=CubicFit())
  # After
  cubic_interp(xs, Series(matrix))

• Saner defaults — less boilerplate: CubicFit() is now the default BC and AutoSearch() is the default search policy (automatically selects the optimal algorithm internally). Explicit bc=CubicFit() and search=LinearBinarySearch() are no longer needed:

  # Before
  cubic_interp(xs, fs; bc=CubicFit(), search=LinearBinary(), extrap=ExtendExtrap())
  # After
  cubic_interp(xs, fs; extrap=ExtendExtrap())

• Derivative dispatch: Val((1, 0)) replaced by DerivOp(1, 0) for ND derivatives; integer deriv=1 replaced by deriv=DerivOp(1) for series interpolants.

• Renamed BC: NaturalBC() → ZeroCurvBC().

• Simplified imports: Selective import FastInterpolations: cubic_interp, deriv1, ... replaced with using FastInterpolations — all public symbols are now properly exported.

AdaptiveArrayPools v0.3

• Reduced pool overhead and improved performance.
• Significantly enhanced compile-time and runtime safety checks (escape detection, structural mutation guards, etc.). Further adoption of these new safety features will follow in a subsequent PR.

⚠️ Bug Fix: Pool-backed arrays escaping scope in Vacuum

While upgrading, I discovered a misuse of AdaptiveArrayPools in compute_vacuum_response: arrays allocated via zeros!(pool, ...) (wv, grri, grre, plasma_pts, wall_pts) were being returned to the caller, escaping their @with_pool scope. Once the scope ends, the pool may reclaim and reuse that memory, leading to silent data corruption. Fixed by replacing with standard zeros() allocations.

AdaptiveArrayPools v0.3 now includes compile-time escape detection that catches most of these patterns automatically at macro-expansion time. Enabling RUNTIME_CHECK mode for test suites and adopting the related safety skills will be addressed in a follow-up PR.

[logan-nc]

@mgyoo86 I am out Th,F then traveling to KSTAR for experiments the next two weeks, so forgive me if I am slow to review this - it is not for lack of appreciation!!

@jhalpern30 please help me out with reviewing this

[logan-nc]

@claude review

[claude]

Code review posted — see below for details.

[claude]

Placeholder — deleting this, real review below.

[claude]

Code Review

Summary: This PR correctly migrates FastInterpolations v0.4 and AdaptiveArrayPools v0.3, and fixes a genuine memory-safety bug in Vacuum.jl. However, several Series() wrappers for multi-column interpolants were missed, which will cause runtime errors on those code paths.

Bug 1 - InverseEquilibrium.jl (~line 304, ffac-normalization branch)
The two other sq constructions in the function use Series(sq_fs), but this one was missed: cubic_interp(sq_xs, sq_fs; extrap=ExtendExtrap()) needs Series(sq_fs).

Bug 2 - ForceFreeStates/Free.jl (~line 158)
wv_flat is a matrix: cubic_interp(psi_array, wv_flat; extrap=ExtendExtrap()) needs Series(wv_flat).

Bug 3 - ForceFreeStates/Bal.jl (~lines 292, 372)
Both hcat(1 ./ bsq, jac .* b1 ./ bsq) and spl1_fs are matrices passed without Series(). The deriv1() calls on these will also fail. These code paths have no test coverage so the failures will only appear at runtime.

Separatrix refactor: hardcoded brackets
Brent avoids Newton divergence but requires a valid bracket. The brackets are hardcoded (outboard midplane at theta near 0, inboard near 0.5, etc.) and Brent will throw an error rather than gracefully failing if the root is outside the bracket — e.g. for a highly triangular equilibrium. A descriptive error message wrapping find_zero would help diagnose this. The midplane find_zero calls also lack atol/rtol that the Z-extremum calls pass — these should be consistent.

Positive notes

• Pool-escape fix in Vacuum.jl is correct and important.
• @views additions and itp_2d_opts deduplication are good.
• The new separatrix/kappa tests are a useful addition.

[jhalpern30]

@mgyoo86 looks good, thanks for doing this! Let me know what you think of Claude's review comments regarding the missing Series and the separatrix Brent solve. Once those are resolved, I can merge.

[mgyoo86]

@jhalpern30
I've addressed the Claude's findings:

• added the missing Series() wrappers for matrix interpolants in Bal.jl, Free.jl, and InverseEquilibrium.jl
• improved the separatrix solver with consistent atol/rtol tolerances and descriptive error messages.

I didn't widen the bracket values, however, the current brackets already span a full half-period each, which should be sufficient and safe to find solutions.

Thanks!

@logan-nc have a safe trip! :)