EQUIL - BUG FIX - Initialize sol_run 4th profile column (cross-run auto-mpsi corruption)

[logan-nc]

Summary

Fixes a cross-run state corruption discovered while building the kinetic regression decks: calling main() repeatedly in one Julia process made the log_asymptotic auto-mpsi for the Solovev equilibrium produce wildly varying grids (mpsi 16 → 600–900, occasionally InexactError), only after a prior kinetic run in the same process. Fresh subprocesses (regression harness, CLI) are unaffected, so it was latent — but test/runtests_fullruns.jl calls main() 6× in one process.

Root cause

sol_run (AnalyticEquilibrium.jl) allocated the source-profile array sqfs = Array{Float64}(undef, ma+1, 4) but filled only columns 1–3, leaving column 4 uninitialized. In a fresh process that memory reads ≈0 (benign); after a kinetic run it holds stale data. The auto-mpsi curvature estimate (_estimate_mid_spacing) reads that 4th column, so the garbage spiked the estimated profile curvature → collapsed the mid-region knot spacing → inflated mpsi.

It is not a pool re-entrancy bug — the AdaptiveArrayPools task-local pool depth returns to 1 after every run (no @with_pool leak). This was plain use-of-uninitialized-memory.

Diagnosed by deterministic reproduction + bisection (a non-kinetic ideal full run does not corrupt; a kinetic full run does) and by instrumenting the per-profile second-derivative estimate (the spike was isolated to column 4).

Fixes

• sol_run: allocate sqfs with zeros so column 4 is a deterministic 0.0, matching the benign fresh-process state the validated baselines were built on.
• _estimate_mid_spacing: size the evaluation buffer to the actual profile count (size(sq_in.y, 2)) instead of a hardcoded 4, so it never reads beyond the real profiles — hardening the 3-profile (F, P, q) builders against the same class of bug. Mirrors InverseEquilibrium.jl, which already sizes this way.

Validation

• Reproduction: kinetic-run-then-auto-mpsi ×4 in one process → before: mpsi 16, 934, 687, 632; after: stable 16 every iteration.
• Full test suite (test/runtests.jl): green.
• Regression harness (--refs develop,local): diiid_n1, solovev_n1, solovev_multi_n, solovev_kinetic_calculated all 0.0-diff (zero numerical impact — confirms column 4 is unused/zero-valued in the validated path).

🤖 Generated with Claude Code