ForceFreeStates - BUG FIX - Pass direction arg to IntegrationChunk in kinetic EL crossing

[logan-nc]

Summary

cross_kinetic_singular_surf! constructed its placeholder IntegrationChunk with a stale 4-argument call:

IntegrationChunk(0.0, 0.0, false, 0) # (psi_start, psi_end, needs_crossing, ising)

but IntegrationChunk is an @kwdef struct whose direction field (added later for bidirectional parallel FM) makes the positional constructor require 5 arguments. The ising=0/direction=1 defaults apply only to the keyword constructor, not the positional one, so the 4-arg call throws:

MethodError: no method matching IntegrationChunk(::Float64, ::Float64, ::Bool, ::Int64)

The fix passes direction=1, matching the two sibling placeholder call sites (cross_ideal_singular_surf! and the Riccati crossing), which already use the 5-arg form.

Why it was latent

The line is only reached on the serial Euler–Lagrange shooting solver (use_parallel=false) when a kinetic singular surface is present (kinetic_factor>0, self-consistent Mode A). Until now, self-consistent kinetic runs always routed through the parallel-FM/Riccati path, which bails out earlier with "kinetic_factor > 0 not implemented yet in Riccati". So this call site was never executed and the defect stayed hidden.

Verification

• Empirically reproduced the MethodError for the 4-arg call and confirmed the 5-arg call succeeds.
• With the fix, a self-consistent kinetic run (use_parallel=false, kinetic_factor=1) finds and crosses the kinetic singular surface and completes end-to-end with converged output.
• Regression harness (diiid_n1, solovev_n1): no attributable change — the touched path is unreachable in these ideal/perturbative cases (confirmed by isolation; the only residual scatter is pre-existing threaded-reduction non-determinism in near-singular PE diagnostics).

🤖 Generated with Claude Code

https://claude.ai/code/session_01SsLUydP2gJbE1tGoaHDiS1

[logan-nc]

@claude review this

[claude]

Review

Correct, minimal fix. IntegrationChunk is @kwdef with a 5th direction::Int = 1 field added later; the positional constructor at EulerLagrange.jl:679 was never updated, so it would MethodError the first time cross_kinetic_singular_surf! was actually reached (serial use_parallel=false + kinetic_factor>0). The 5-arg form now matches the two sibling call sites (cross_ideal_singular_surf!:610, Riccati.jl:1150).

A couple of observations, not blocking:

• Minor inconsistency: the sibling call at EulerLagrange.jl:610 passes the real ising value as the 4th positional arg, while this fix keeps 0 (IntegrationChunk(0.0, 0.0, false, 0, 1)) instead of the ising::Int argument already in scope at line 670/676. I confirmed sing_der! (Sing.jl:1105) discards the chunk (ctrl, equil, ffit, intr, odet, _ = params), so this is functionally inert either way — but passing the real ising would be more consistent with the sibling call and less surprising if the chunk is ever consumed there in the future.
• No regression/unit coverage added: per the PR description, this path (use_parallel=false with kinetic_factor>0) isn't exercised by any example config or test (use_parallel=false doesn't appear in any examples/*/gpec.toml, and there's no test referencing cross_kinetic_singular_surf!). That's presumably why this went undetected for as long as it did. Worth adding a small test/example exercising this branch (or at least a targeted unit test constructing the args and calling cross_kinetic_singular_surf!) so a future struct-field change can't silently reintroduce the same class of bug — the regression harness alone won't catch it since none of its cases hit this path.
• Per CLAUDE.md, this PR should still go through regression-guardian before merge even though the description already reports "no attributable change" — worth confirming that was run against current develop, not a stale baseline.

Otherwise the change is exactly scoped (1 line) and matches the existing pattern used elsewhere in the file.

[logan-nc]

Addressed the review:

• ising consistency (point 1): amended to IntegrationChunk(0.0, 0.0, false, ising, 1), passing the in-scope ising to match the sibling call sites. Still functionally inert (the chunk is discarded in sing_der!), but consistent.
• regression-guardian (point 3): run against current develop (not a stale baseline) — diiid_n1, which is the case that exercises the kinetic FGAR quantities. No attributable change:NTV torque FGAR and dW FGAR are OK, and the rows that show a delta (delta prime ~4.5%, resonant field ~4.6–6.6%, island widths ~2.5%) are the known threaded-reduction non-determinism in near-singular PE diagnostics — identical to the develop-vs-develop baseline, not caused by this change. (The ideal diiid_n1 path never reaches cross_kinetic_singular_surf!, so the fix can't affect it.)
• Test coverage (point 2): deferring — this is a runtime MethodError on a slow self-consistent path, so a tracked regression case would be noise rather than signal. Noted for a future lightweight guard if the placeholder call sites are revisited.